International Journal of Multiphase Flow 76 (2015) 122–143

Contents lists available at ScienceDirect

International Journal of Multiphase Flow

j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / i j m u l fl o w

Multiscale simulation of atomization with small droplets represented

by a Lagrangian point-particle model
Y. Ling a,b,⇑, S. Zaleski a,b, R. Scardovelli c
a
Institut Jean Le Rond d’Alembert, Sorbonne Universités, UPMC Univ Paris 06, UMR 7190, F-75005 Paris, France
b
Institut Jean Le Rond d’Alembert, CNRS, UMR 7190, F-75005 Paris, France
c
DIN – Lab. di Montecuccolino, Università di Bologna, Via dei Colli 16, 40136 Bologna, Italy

a r t i c l e i n f o a b s t r a c t

Article history: Modeling and simulation of atomization is challenging due to the existence of a wide range of length
Received 13 January 2015 scales. This multiscale nature of atomization introduces a fundamental challenge to numerical simula-
Received in revised form 1 June 2015 tion. A pathway to comprehensive modeling is still to be found. The present study proposes a multiscale
Accepted 9 July 2015
multiphase flow model for atomization simulations, where the large-scale interfaces are resolved by the
Available online 17 July 2015
Volume-of-Fluid (VOF) method and the small droplets by the Lagrangian point-particle (LPP) model.
Particular attention is focused on the momentum coupling between LPP and resolved flow and the con-
Keywords:
version between droplets represented by VOF and LPP. A series of multiphase flow problems are consid-
Multiphase flows
Atomization
ered to validate the model. The results obtained by a number of simulations are compared against direct
Lagrangian point-particle numerical simulation (DNS) results and experimental data. In particular, the model is applied to simulate
Volume of fluid method the gas-assisted atomization experiment, and the numerical results are compared to the experimental
Multiscale modeling measurements for a quantitative validation.
Ó 2015 Elsevier Ltd. All rights reserved.

Introduction In the past decades, sophisticated numerical schemes have been

Atomization is an important fundamental multiphase flow
problem. As it is essential to many engineering applications, such
as fuel injection in combustion engines, many experimental and
numerical studies on the topic have been reported in the literature
(Lasheras et al., 1998; Marmottant and Villermaux, 2004;
Descamps et al., 2008; Fuster et al., 2009; Shinjo and Umemura,
2010; Tomar et al., 2010; Matas et al., 2011; Fuster et al., 2013).
Nevertheless, the mechanisms that control spray formation and
evolution are far from being understood. Numerical simulation
has been shown to be a powerful tool to investigate atomization,
in particular in providing physical insight into fundamental phy-
sics and aspects of the phenomenon that are difficult to be exam-
ined experimentally. However, modeling and simulation of
atomization is challenging due to the existence of a wide range
of length scales, as in an atomizing liquid jet they can vary from
the length of the jet (several cm) to the size of the smallest droplets
(submicrons). This multiscale nature of atomization introduces a
fundamental challenge to numerical simulation. A pathway to
comprehensive modeling is still to be found.
⇑ Corresponding author at: Institut Jean Le Rond d’Alembert, Université Pierre et
Marie Curie, 4 Place Jussieu, 75252 Paris, France.
E-mail address: yueling@dalembert.upmc.fr (Y. Ling).
developed to capture or track the interface evolution, such as the
Volume-of-Fluid (VOF) (Hirt and Nichols, 1981), Level-Set (LS)
(Sussman et al., 1994), and Front-Tracking methods (Tryggvason
et al., 2001). With these numerical schemes accurate direct numer-
ical simulations (DNS) of multiphase flows with complex interfaces
can be conducted (Fuster et al., 2009; Herrmann, 2010; Shinjo and
Umemura, 2010; Fuster et al., 2013). In particular, the VOF method
has the advantage of conserving mass and is in a good balance
between accuracy and implementation complexity. In previous
studies of our group, VOF has been applied to a variety of multi-
phase flows with complex interface topology change, like primary
atomization, droplet splash, rising bubble, yielding accurate simu-
lation results (Josserand et al., 2005; Fuster et al., 2009; Popinet,
2009; Fuster et al., 2013).
Direct numerical simulation with interface capturing/tracking
schemes is a very powerful tool to understand the fundamental
physics of atomization. However, the range of scales that can be
covered by DNS with today’s computing capability is still smaller
than what would be needed for practical applications or even com-
parison to experiments. Adaptive Mesh Refinement (AMR) tech-
niques have been developed and utilized in multiphase flow
simulation to reduce the cost of computation (Popinet, 2009;
Agbaglah et al., 2011). In AMR, a fine mesh is used only in the
regions that require a higher resolution, such as near the interface

http://dx.doi.org/10.1016/j.ijmultiphaseflow.2015.07.002
0301-9322/Ó 2015 Elsevier Ltd. All rights reserved.
 Y. Ling et al. / International Journal of Multiphase Flow 76 (2015) 122–143
and in shear layers. Nevertheless, even with AMR, DNS is still often
too expensive for many atomization problems. First of all, it is well
known that octtree AMR adds a large overhead to computations.
This is mainly due to the placement of objects in memory, resulting
numerous page faults. Because 3D block-AMR requires the creation
of a very large number of boxes, similar difficulties may also apply
in that case. Second, the parallelization of AMR adds additional
challenges. Regions in which a high density of small computations
cells exist are not always distributed uniformly in space (in
atomization they are close to the nozzle) which makes domain
decomposition into blocks of equal size difficult. Third, the most
demanding mesh requirement is usually introduced by the small
droplets that are formed in atomization. Typically, to obtain rea-
sonable results of motion and interface dynamics for a spherical
droplet, at least eight to twelve grid points across the droplet
diameter are required. It is known from experiments that the
diameter of the atomizing liquid jet is about three orders of mag-
nitude larger than the smallest droplets. For such a wide range of
length scales, the number of grid points that is required to resolve
every scale, namely a true DNS, is extremely expensive even with
AMR.
An alternative approach to resolve this multiscale challenge in
atomization simulation is to introduce subgrid models, although
the two ideas could be combined. The idea of subgrid modeling
is very simple: only the physical scales larger than the cell size
are resolved, while for the physical scales that are smaller suitable
models are introduced to represent the relevant physics. A good
example of subgrid modeling is the Large-Eddy Simulation (LES)
model for turbulence (Pope, 2000). In LES, a low-pass filtering is
applied and only the large eddies are resolved. As a result, a coarser
mesh compared to DNS of turbulence can be used. However, to
accurately capture the large-scale physics, the small unresolved
turbulent eddies and their interaction with the large scales have
to be represented by a subgrid model, such as the Smagorinsky
model (Smagorinsky, 1963). For turbulent multiphase flows like
atomization, subgrid modeling is more challenging since both the
unresolved turbulence and multiphase scales, and the interaction
between two unresolved scales need to be modeled properly
(Lakehal and Liovic, 2011; Lakehal et al., 2012; Chesnel et al.,
2011a,b; Vincent et al., 2010; Larocque et al., 2010; Toutant
et al., 2009a,b; Labourasse et al., 2007; Sirignano, 2005). If a spatial
filter of a constant filtering length is applied to the whole flow field,
then both the flow properties and the phase-characteristic function
need to be filtered (Lakehal and Liovic, 2011; Sirignano, 2005). As a
result, an ideal subgrid model would need to capture a variety of
multiphase dynamics, such as the formation of sheets, ligaments,
and droplets, droplet dynamics, droplet impact on liquid interface,
and the effect of both resolved and unresolved turbulence on these
pheomena. Although progress has been made in the past decade
(Lakehal et al., 2012), such comprehensive subgrid models are
exceedingly complex and the current knowledge of many of the
above phenomena still requires significant improvement for the
development of accurate models.
Nevertheless, as the droplets formed in atomization are usually
very small, it is easier to only apply subgrid modeling for their
motion, and the Lagrangian point-particle (LPP) approach is a per-
fect candidate. The LPP model has been widely utilized in
particle-laden flows and sprays (Balachandar and Eaton, 2010;
Apte et al., 2003). In LPP, the particles (or droplets) are viewed as
point masses and the flow around individual particles is not
resolved. To accurately track the particles and the backward effect
to the carrier flow, physical models are needed to calculate the
instantaneous force acting on the particles. Eventually, in a
so-called Eulerian–Lagrangian simulation, the macro-scale fluid
flow is resolved in an Eulerian framework and the point-particles
123
are evolved in a Lagrangian framework. In atomization, the typical
Weber number of the droplets, Wep , is usually very small. (Precise
definition of Wep will be given in Section ‘Modeling and numerical
methods’.) As a result, the droplets remain close to a spherical
shape, and the LPP model is suitable to capture their motion and
the backward effect to the flow. From the filtering perspective,
LPP can be viewed as a local spatial filter with a variable filtering
length scale (Climent and Magnaudet, 2006). Since it is difficult
to filter the overall flow field, filtering is only applied at the particle
location with a filtering length scale which is related to the droplet
size. In this work, a multiscale simulation approach that couples
the LPP model and an interface-capturing scheme is proposed. In
this approach, the goal is to filter the scales on the order of a small
number, say nc , of grid cells, and to represent larger scales. Thus
droplets of a diameter smaller than or of the order of nc Dx are rep-
resented by the LPP model, while larger droplets and fluid masses
are represented by an interface tracking scheme. Thus part of the
flow is represented in a DNS-like manner while both the flow
inside the small droplets and the surrounding flow perturbed by
these droplets are filtered. Thus the motion of the small droplets
and their backward effect to the resolved flow are represented by
the LPP force model. Moreover, care is taken to work with
Reynolds and Weber numbers sufficiently moderate to minimize
the energy present at such small scales.
In recent years, several pioneering efforts have been made to
couple the LPP model with interface-capturing schemes, such as
LS (Herrmann, 2010) and VOF (Tomar et al., 2010). The results
are quite encouraging and show the great potential of this multi-
scale approach for accurate large-scale simulations of atomization.
However, there still exist unresolved issues in the coupling of LPP
with interface-capturing schemes which require further investiga-
tion and validation. In conventional simulations with the LPP
model the grid cell is much larger than the particle size, typically
dp < Dx=10, where dp and Dx represent the particle diameter and
cell size, respectively, as shown in Fig. 1(a). As a result, a single
computational cell may contain many particles. Sometimes, the
number of particles in a cell is so large that a super-particle tech-
nique is utilized, in which one computational particle (or parcel)
is used to represent many physical particles. In contrast, for a fully
resolved droplet (RD), typically about ten grid points per diameter
are required, as in Fig. 1(d). The ratio dp =Dx has a significant jump
of about two orders of magnitude between these two limits. Even if
AMR is considered, it is impossible to adapt the mesh to such a spa-
tial jump in a single time step when a resolved droplet is converted
to a Lagrangian particle or vice versa. In yet another numerical
setup, if the grid were stretched and the droplet moved from a high
density grid to a coarse grid region, a transition region of poor res-
olution would again be encountered. Thus in all cases the mesh has
to be adapted gradually, covering regimes where the grid resolu-
tion of the droplet lies in between the two extreme limits, as in
Fig. 1(b) and (c). For these two intermediate conditions, dp is only
a few times larger than Dx, and with such a resolution the flow
inside and outside the droplet is not captured accurately by DNS.
It is therefore reasonable to convert these poorly-resolved droplets
to LPP, but because of the larger-than-grid-size droplets the
required modeling is much more challenging than in conventional
case.
The challenge mainly comes from the coupling between the LPP
droplets and the resolved fluid flow. The fluid-dynamic force acting
on each individual droplet is crucial to LPP models, since it deter-
mines both the droplet motion and the backward effect to the fluid.
Typically, the calculation of this force requires the undisturbed
fluid velocity at the droplet location. For the simple case of a single
droplet, the ‘‘undisturbed fluid velocity’’ can be considered as the
fluid velocity without the local perturbation induced by the
124 Y. Ling et al. / International Journal of Multiphase Flow 76 (2015) 122–143
Fig. 1. Droplet representation at different grid resolutions: (a) dp  Dx=10, (b) dp  Dx, (c) dp  4 Dx, and (d) dp  10 Dx.
droplet, or the fluid velocity far away from the droplet location if a
uniform flow is imposed. For the case with many droplets, the
undisturbed fluid velocity ‘‘seen’’ by an LPP droplet can be consid-
ered as the fluid velocity when the droplet of interest is absent but
all the other droplets are present (Balachandar, 2009). Because the
droplets and the fluid are two-way coupled, not only the fluid
exerts a fluid-dynamic force on the droplets, but this force is also
exerted back to the fluid with an opposite sign.
Conventionally, the droplet force is applied to the fluid as a
point source located at the center of mass of the droplet (Climent
and Magnaudet, 2006). Strictly speaking, after the droplet force is
added, the undisturbed fluid velocity is not available anymore,
since the fluid flow has been ‘‘perturbed’’. However, for a droplet
that is much smaller than the computational cell the perturbation
induced by each individual droplet to the fluid within a cell is
indeed small. As a result, the fluid velocity stored in the cell is a
good approximation of the undisturbed fluid velocity. On the other
hand, for droplets that are larger than the cell size, if the force is
also applied as a point source then the fluid velocity of the cell,
where the center of mass of the droplet is located, will be influ-
enced significantly and it cannot be considered equal to the undis-
turbed fluid velocity. Furthermore, if the undisturbed fluid velocity
is not computed correctly, an error will be present in the droplet
force and it will propagate back to the resolved fluid flow, thus
leading to serious computational errors (Balachandar, 2009;
Eaton, 2009).
It should be reminded that, even though the local perturbation
of the flow at the scale of the droplet size is ignored in the LPP
model, the LPP droplets and the resolved fluid flow are still
two-way coupled. Since the force exerted on each individual LPP
droplet is subtracted from the resolved fluid flow, the presence
of LPP droplets will influence the resolved flow. However, this
influence is only at a spatial scale that is much larger than the dro-
plet size (Balachandar, 2009).
In the conversion from a resolved droplet to an LPP droplet
(RD-to-LPP) the phase-characteristic function v in the region occu-
pied by the droplet is changed from the value v ¼ 1 of the liquid
phase to v ¼ 0 of the gas, as in Fig. 2. Moreover, the local perturba-
tion induced by the droplet in the fluid velocity field should be
removed carefully, as depicted in Fig. 2, so that the ‘‘undisturbed
fluid velocity’’ is available to compute the fluid-dynamic force act-
ing on the LPP droplet. The inverse conversion (LPP-to-RD) requires
a similar treatment.
In recent years, more sophisticated LPP models have been pro-
posed for larger-than-grid-size solid particles, such as the
Force-Coupling method (FCM) by Maxey and his colleagues
(Maxey et al., 1997; Maxey and Patel, 2001; Lomholt et al., 2002;
Lomholt and Maxey, 2003). Instead of computing the actual
fluid-dynamic force exerted on the particle, a Gaussian distribution
of external force monopoles and dipoles is applied to mimic the
local perturbation. In such a case, the undisturbed fluid velocity
is not required to model the interphase coupling. However, FCM
is developed based on the multipole expansion of Stokes flows.
Extension to droplets with finite Reynolds number is difficult.
u χ=0 u χ=0
χ=1 up χ=0
up
Resolved Drop Lag Point-Particle
Fig. 2. Schematic representation of a resolved droplet and a Lagrangian point-
particle droplet.
Even for droplets with small Reynolds number, FCM typically
requires four to six grid points per particle diameter (Lomholt
and Maxey, 2003). It is thus inapplicable in atomization simula-
tions, since we need to deal with droplet resolution dp =Dx going
continuously from the very small to four grid points.
The goal of the present study is to propose a multiphase flow
model for atomization simulations, where the large-scale inter-
faces are resolved by the VOF method and the small droplets by
the LPP model. To focus on the development and validation of
the LPP model, in particular for droplets that are larger than the
grid size, we perform the study on a fixed grid. The application
of the LPP model in an octree AMR framework will be considered
in future works. When an adaptive grid is used, the cell size can
be increased (gradually) by at least an order of magnitude, after
the small resolved droplets are converted to LPP droplets. In such
a case, the total computational cost can be reduced significantly.
For a fixed grid, even though the reduction of computational time
is not striking, there are several other important advantages to rep-
resent small poorly resolved droplets by the LPP model. First, the
LPP model typically yields more accurate results on droplet
dynamics than keeping the droplets resolved by VOF with poor res-
olution (dp < 4Dx). Second, the poorly resolved droplets are usually
removed in VOF simulations (Shinjo and Umemura, 2010). An
unaccounted-for removal of the small droplets formed makes the
measurements of probability density function (PDF) of droplets
impossible and also leads to a loss of physics. With the LPP model,
these droplets can be kept as LPP droplets. Third, the
poorly-resolved droplets sometimes cause numerical instability
(which is the motivation to remove them), converting them to
LPP improves the numerical stability of the simulation and speeds
up the convergence of Poisson solver slightly.
The multiphase flow models, corresponding equations and
numerical methods are presented in Section ‘Modeling and numer-
ical methods’. A series of multiphase flow problems are considered
to validate the approach, and the results obtained by a number of
simulations are compared against DNS results and experimental
data in Section ‘Results’. In particular, the model is applied to sim-
ulate the gas-assisted atomization experiment by Descamps et al.
(2008). This experiment is of particular interest because the drop
trajectories and drop flying angle statistics are measured. The
 Y. Ling et al. / International Journal of Multiphase Flow 76 (2015) 122–143 125

numerical results are compared to the experimental measure-

Start
ments for a quantitative validation. Finally, we summarize the con-
clusions of the present study in Section ‘Conclusions’.

Modeling and numerical methods NS Solver RF-LPP

Coupling LPP Solver
Our numerical approach is made up of two main components. Color
The first one for the resolved flow contains a VOF method similar Function
to that implemented in several open-source codes such as Solver (VOF)
SURFER (Lafaurie et al., 1994), Gerris (Popinet, 2003) and Basilisk
LPP-to-RD No Qualify
(Popinet). The other one is the Lagrangian point-particle (LPP) Tag Drops Conversion LPP?
method, which is the main topic of this paper. The combined
approach has been implemented in the more recent free code Yes
PARIS-Simulator (Arrufat et al.). Qualify Yes RD-to-LPP
The LPP model can be viewed as a local spatial filtering on top of LPP? Conversion
the small droplets with a variable filtering length scale. The LPP
No
droplets are those with a diameter smaller than the cut-off droplet
size in the flow field. The cut-off droplet size is generally a function
of the grid resolution, say dcut  ð4—6Þ Dx, since droplets with a No
t>t-End?
diameter smaller than dcut are poorly resolved. Furthermore, dro-
plets should also satisfy the physical requirement of an associated Yes
RF: Resolved Flow
small Weber number, Wep , so that they remain almost spherical. RD: Resolved Drop
LPP: Lagrangian Point-Particle End
The filtered flow contains both phases and is governed by the
incompressible, variable-density, Navier–Stokes equations, with Fig. 3. Flow chart of the combined VOF-LPP algorithm.
the interface evolution described by the VOF method. However,
the LPP model and the coupling techniques here presented can
be applied to other interface methods, such as Level-Set and @ t C þ u  rC ¼ 0: ð3Þ
Front-Tracking. The fluid flow around and inside the droplets is
The fluid density and viscosity are then defined by
not resolved, and the LPP force model is used both to calculate
the droplet motion and as a closure model to account for the effects q ¼ C ql þ ð1  CÞqg ; ð4Þ
of the droplet dynamics on the resolved or filtered flow field.
The flowchart of the combined algorithm is shown in Fig. 3. For
l1 ¼ C l1
l þ ð1  CÞl1
g or l ¼ C ll þ ð1  CÞlg ; ð5Þ
the resolved flow, the Navier–Stokes (NS) equations and the color where the variables with subscripts ‘‘g’’ and ‘‘l’’ represent properties
function advection equation are first advanced in time. Then the of gas and liquid, respectively. The variables without a subscript are
separated liquid structures are identified and tagged by examining corresponding to the VOF representation, which denote variables of
the volume fraction C field. The resolved droplets that are qualified gas, liquid, and gas–liquid mixture for C equal to 0, 1, and a frac-
for LPP are then submitted to the RD-to-LPP conversion routine. On tional number, respectively.
the other side, the equation of motion for each LPP droplet is solved The arithmetic mean is used for density, while the arithmetic or
to update its velocity and position. The LPP droplets are then exam- harmonic means are used for viscosity according to the physical
ined to check if they are still qualified to remain particles, if not problem (Boeck et al., 2007; Vincent et al., 2014). Furthermore,
they are converted back to resolved droplets. While the resolved the present approach assumes that the velocity of the two phases
flow properties are needed to integrate the LPP equation of motion, match at the interface, which is valid for incompressible viscous
the force exerted by the flow on the LPP droplets should be sub- fluids without phase change (Tryggvason et al., 2011; Rudman,
tracted from the flow momentum equation. Thus the NS and LPP 1998; Le Chenadec and Pitsch, 2013).
solvers are two-way coupled. A description of the various solvers The third term on the right hand side of Eq. (1) is a singular
and routines is given in the following sections. term, with a Dirac distribution function ds localized on the inter-
face, and it represents the surface tension force. The surface ten-
Governing equations and numerical methods for the resolved flow sion coefficient is r, and j and n are the local curvature and unit
normal of the interface.
Governing equations Finally, the last term of Eq. (1), f p , is the closure term that
The one-fluid approach is employed to compute the resolved accounts for the backward effect of the LPP droplets on the
two-phase flow, where the liquid and gas phases are treated as resolved flow.
one fluid with material properties that change abruptly across
the interface. The incompressible, variable-density, Navier–Stokes
Numerical methods
equations with surface tension are
Navier–Stokes solver. The Navier–Stokes equations, Eqs. (1) and (2),
qð@ t u þ u  ruÞ ¼ rp þ r  ð2lDÞ þ rjds n  f p ; ð1Þ are solved by the projection method (Chorin, 1968). The time inte-
r  u ¼ 0; ð2Þ gration is conducted by a second-order predictor–corrector
method (Tryggvason et al., 2011). The finite-volume approach on
where q and l are the fluid density and viscosity, u and p the veloc- a regular, cubic staggered grid is utilized for spatial discretization.
ity and pressure fields, and D the deformation tensor with compo- The advection term is discretized by the third-order QUICK scheme
nents Dij ¼ ð@ i uj þ @ j ui Þ=2. (Leonard, 1979). The viscous term is treated explicitly with the
The volume fraction C is introduced to distinguish the different standard second-order centered difference scheme. In the projec-
phases, in particular C ¼ 1 in the computational cells with only the tion step, an elliptic equation for the pressure is solved by a red–
liquid phase, and its time evolution satisfies the advection black successive overrelaxation (SOR) Gauss–Seidel iteration
equation method, which is found to be competitive with a more complex
126 Y. Ling et al. / International Journal of Multiphase Flow 76 (2015) 122–143

 
multigrid solver, probably because of the small time step dup u ~ g  up qg Du~ g C m qg Du~ g dup
¼ /þ þ  þ g; ð6Þ
constraint. dt sp qp Dt qp Dt dt

where up is the particle velocity and u ~ g the local undisturbed gas

Volume-of-Fluid method. The piecewise-linear geometrical
Volume-of-Fluid method is used to solve the color function advec-
tion equation Eq. (3) (Scardovelli and Zaleski, 1999). The overall
VOF method proceeds in two steps: interface reconstruction and
then its advection. The interface reconstruction is performed using
a piecewise-planar interface representation in each cell. The inter-
face normal is computed by the Mixed-Youngs-Centered (MYC)
method (Aulisa et al., 2007). After the reconstruction,
direction-split geometrical fluxes are computed to advect the
interface (DeBar, 1974; Rudman, 1997). The VOF method has been
shown to preserve sharp interfaces and be close to second-order
accuracy for practical applications (Aulisa et al., 2007; Popinet,
2009; Bornia et al., 2011).
Curvature and balanced-force surface tension calculation. The
height-function (HF) method is employed to calculate the local
interface curvature, in a manner similar but not identical to the
method implemented in Gerris (Popinet, 2009). We compute
height function values whenever a 1D block of cells exists with a
full cell on one side and an empty one on the other side. The block
should be at most of length N H ¼ 9 cells. We then progress with a
hierarchical sequence of steps: (i) when all the height values in a
3  3 stencil that includes the cell under examination are present,
the curvature is computed by a finite difference calculation
(Popinet, 2009); (ii) when only 7 or 8 heights are found, the result-
ing interface points are fitted by an elliptic paraboloid that approx-
imates the shape of the local interface. However, the height values
must be of the same kind, which means no mixing of the heights
along the z direction, with widths or depths along x and y. The
derivatives of the height function are then given in terms of the
coefficients of the paraboloid; (iii) when less than 7 or 8 heights
are found, we use a mixture of heights, widths and depths to get
a set of at least 7 points. This set is trimmed to eliminate points
that are too close to each other: when two points are less than half
a cell size apart, only one of the two points is kept. Unlike (Popinet,
2009), the coordinate system is not rotated and the points are
directly fitted to an elliptic paraboloid with the axis along the
direction with the largest normal component; (iv) when the previ-
ous option fails, all heights, widths and depths are discarded and
we use the centroid of the interface fragments reconstructed with
the VOF method to create a set of points, which are then fitted to
an elliptic paraboloid.
As in several other works (Renardy and Renardy, 2002; Francois
et al., 2006; Popinet, 2009), a balanced-force surface tension
discretization is employed. We have checked that the method
produces accurate curvatures in 2D and 3D test cases and
well–behaved droplet oscillations in air–water systems.
Lagrangian point-particle model for the unresolved small droplets
The Lagrangian point-particle approach has been widely used
in turbulent dispersed multiphase flows (Balachandar, 2009;
Balachandar and Eaton, 2010). The key idea of the LPP model
is to approximate the dispersed phase, made up of small solid
particles, droplets or bubbles, as point masses. As a result, the
flow at the scale of these particles is not resolved. To accurately
track the LPP droplets in the Lagrangian framework, the force
exerted on each individual droplet is calculated in terms of
the undisturbed flow field properties. The closure is typically
given by the force model or the so–called equation of motion
(EOM) (Crowe et al., 1998; Maxey and Riley, 1983; Gatignol,
1983)
velocity.
The variables with a subscript p are corresponding to the LPP
droplets. In the present study the LPP model is only used to repre-
sent liquid droplets, therefore physical properties denoted by the
subscript p correspond to those of the liquid phase, for example
qp ¼ ql . Furthermore, the liquid volume fraction is identical to
zero, i.e., C ¼ 0, at the cell where an LPP droplet locates, as the liq-
uid is represented by the LPP model. Therefore, the resolved flow
properties such as u; q, and l at the LPP droplet location are all cor-
responding to the gas phase, i.e., ujx¼xp  ug ; qjx¼xp  qg , and
ljx¼xp  lg .
The response time, density, and added-mass coefficient of the
LPP droplet are denoted by sp ; qp , and C m , respectively. The modi-
fied gravity acceleration is denoted by g ¼ ð1  qg =qp Þg 0 , where g 0
is the gravity acceleration.
The droplet Reynolds and Weber numbers can then be defined
as
qg dp ju~  up j
Rep ¼ ; ð7Þ
lg
qg dp ju~  up j2
Wep ¼ : ð8Þ
r
The position xp of an LPP particle is then updated by integrating
dxp
¼ up : ð9Þ
dt
The force terms on the right hand side of Eq. (6) represent the
quasi-steady force, the pressure-gradient force, the added-mass
force, and the gravity force, respectively, while the Basset-history
force has been ignored for simplicity (Maxey and Riley, 1983;
Ling et al., 2013). The relative importance in turbulent flows of
the unsteady forces compared to the quasi-steady force is investi-
gated by a scaling argument in Ling et al. (2013). The unsteady
forces are necessary to accurately capture the LPP droplet motion
when the LPP droplet size becomes comparable to the
Kolmogorov scale and the density ratio qg =qp is significantly large.
In the present study, we consider droplets with a rather small ratio
qg =qp , therefore it is expected the contribution of the unsteady
forces to be small. Nevertheless, the pressure-gradient and
added-mass forces are retained to capture the leading order effect
of the unsteady mechanisms.
The quasi-steady force in Eq. (6) is written as the Stokes drag
multiplied by a correction function /, which takes into account
the effect of finite Rep . In the case of liquid droplets, for which
ll  lg , the standard drag correlation for solid particle can be
employed (Clift and Gauvin, 1970)
!1
42500
/ðRep Þ ¼ 1 þ 0:15Re0:687
p þ 0:0175Rep 1 þ : ð10Þ
Re1:16
p
The response time sp of the LPP droplet is expressed as (Clift
et al., 1978)
qp d2p
sp ¼ : ð11Þ
18 lg
A second-order predictor–corrector method is used for the time
integration of the particle velocity and position, Eqs. (6)–(9), which
is consistent with the algorithms used for the resolved flow.
In order to focus on the development of the VOF-LPP model,
droplet–droplet interaction (i.e., four-way coupling effect) is
 Y. Ling et al. / International Journal of Multiphase Flow 76 (2015) 122–143
neglected in the present LPP model. For the gas assisted atomiza-
tion configuration of interest here, the small droplets formed in
atomization are dispersed downstream of the liquid jet. As a result,
their total volume fraction is quite low and the inter-droplet inter-
action is of secondary importance. Nevertheless, there exist
atomization applications where the effects of inter-droplet interac-
tion and secondary breakup are important (Apte et al., 2003), it is
interesting to extend the present model to include these effects in
future works.
Momentum coupling between LPP and resolved flow
The force exerted by the resolved flow on the LPP droplets is
referred to as forward coupling and its computation is based on
the resolved flow properties, Eq. (6). Due to Newton’s third law,
the force exerted on the LPP droplets needs to be subtracted from
the resolved flow and is referred to as backward coupling. This
forcing term is represented by f p in the resolved flow momentum
equation, Eq. (1).
In the conventional LPP model, the computational cell is usually
much larger than the LPP droplet size and many droplets are con-
tained in each cell. The forcing term f p is then computed by a vol-
ume average over the LPP droplets
N as an approximation of u
1 X p;c
fp ¼ F fp;i ; ð12Þ
V c i¼1
where V c and N p;c are the cell volume and the number of droplets in
that cell. Note that only the force due to fluid-LPP-droplet interac-
tion needs to be subtracted from the momentum equation
 
dup;i
F fp;i ¼ mp;i g : ð13Þ
dt
Although there exist issues of coupling LPP back to the resolved
flow due to the number density fluctuation (Sundaram and Collins,
1996; Ling et al., 2010), the handling of two-way coupling is rela-
tively easy. Since the fluid velocity corresponds to an average value
over the cell, and the cell contains a large number of LPP droplets,
then their influence is automatically filtered out (but the effects of
the ambient flow and other LPP droplets outside the cell are
retained). As a result, the stored fluid velocity can be used directly
as the undisturbed fluid velocity in Eq. (6).
However, in the present LPP model the droplets are bigger than
the cell size, the two-way coupling becomes more complicated and
in particular Eq. (12) is not valid any more. What’s even worse is
the fact that if the force F fp;i is applied to the cell where the mass
center of the LPP droplet is located, or is distributed to all the cells
occupied by the LPP droplet, the resolved flow will be affected
significantly.
For this reason, we retain Eq. (6) and, as suggested by Maxey
et al. (1997), write the resolved velocity ug as the sum of two con-
tributions, ug ¼ u ~g þ u ^ g . The first term u
~ g represents the undis-
turbed fluid velocity, the second one u ^ g is the velocity
disturbance induced by the droplet under examination, a term
which is filtered out in the conventional LPP model.
When there are only a few LPP droplets in the flow field, their
contribution to the momentum equation of the resolved flow is
small, and a simple solution is to neglect the coupling force term
f p in Eq. (1). This is the one-way coupling approximation for which
^ g ¼ 0 and u
u ~ g ¼ ug . It has been found in previous works that the
effect of the coupling force on changing the fluid momentum is
related to the mass fraction of dispersed phase (Crowe et al.,
1998; Ling et al., 2013), hence in the one-way coupling assumption
the error in u ~ g increases with the mass fraction of the LPP droplets
(Eaton, 2009).
127
~ g , we propose to calcu-
In order to obtain a better estimate of u
late f p as
Np
X
fp ¼ F fp;i Gðx  xp;i Þ; ð14Þ
i¼1
where N p is now the total number of LPP droplets in the flow field.
The Gaussian function Gðx  xp;i Þ is a numerical representation of
the LPP droplet coupling force (Maxey et al., 1997)
3=2
GðxÞ ¼ ð2pL2 Þ expðjxj2 =2L2 Þ; ð15Þ
where L controls the size of the region where the force should be
distributed. Eq. (14) is similar to that given by Maxey et al.
(1997), but the interpretation is different. In the present model, L
is viewed as the length scale of a spatial filter of the coupling force
at the LPP droplet location. The parameter L is chosen to be larger
than the actual droplet size. Depending on the volume fraction of
the LPP droplets, L is picked between 5 and 10 dp . Therefore, the
th
coupling force F fp;i corresponding to the i LPP droplet is smoothly
distributed within a region larger than the droplet size, rather than
concentrated at the droplet location. At every LPP droplet location,
the contribution of the force from the present droplet to the local
resolved flow is small, so that ug computed from Eq. (1) can be used
~ g in Eq. (6).
There are several advantages for treating the coupling force in
the above way. First, the undisturbed fluid velocity is directly given
by the resolved flow momentum equation to calculate the droplet
force. Second, the integration of Eq. (14) over the whole domain
yields the total force exerted on all the LPP droplets and the total
momentum of the combined RF-LPP system is conserved. Third,
since L is determined by the droplet diameter and not by the cell
size, the coupling between RF and LPP droplets does not depend
on the grid. This feature is quite important if the LPP model is cou-
pled with AMR, because the calculation of the force on the LPP dro-
plets and the backward effect on the resolved flow will not be
affected when the mesh adapts itself from the well-resolved dro-
plet regime, Dx < 0:1dp , to the conventional LPP model, Dx > 10dp .
The present treatment of the coupling force will only capture
the backward effect of the LPP droplets on the resolved flow on
length scales larger than L. The dynamics due to fluid-droplet
interaction on scales smaller than L, such as the shear layer near
the droplet interface, the wake of the droplet and the interaction
with the resolved flow, will be missed.
Finally, since the LPP droplets do not in general coincide with
the locations at which the fluid velocity is stored, interpolation is
required to compute ug at the LPP droplet location, and a
tri-linear interpolation scheme is used in the present study (Apte
et al., 2003; Ling et al., 2010). When the spatial variation of the
resolved flows is significant over the LPP droplet size, then picking
the fluid velocity at the LPP droplet center is not sufficient and the
Faxén correction is often used to generalize the force models
(Faxén, 1922). Furthermore, if the LPP droplet moves in a shear
flow, then the shear lift force will be induced (Legendre and
Magnaudet, 1997). Nevertheless, the LPP model is only applied to
small droplets in the present study, the effect of inhomogeneous
ambient flows on the droplet force is expected to be generally
small and thus the Faxén and lift forces are neglected.
Two-way conversion between LPP and resolved droplet
The flow chart of the combined algorithm in Fig. 3, shows the
tight coupling between the two solvers for the resolved flow and
the LPP droplets. In atomization simulations that consider the jet
formation as well, the initial flow is completely resolved with the
VOF method, since small droplets are not yet formed. Once the
128 Y. Ling et al. / International Journal of Multiphase Flow 76 (2015) 122–143

liquid jet starts to atomize, the droplets away from the interface for icell =1,Ncell do
are converted to LPP droplets. Some of the LPP droplets may follow Identify separated droplets and generate RD array
Tagging
trajectories that bring them back toward the interface. These dro- Perform droplet calculation (droplet volume, etc)
end for
plets may be converted back to resolved droplets. The conversion
for iRD =1,NRD do
algorithm is summarized in Fig. 4.
Check conversion criteria
if Droplet iRD qualified for LPP
Conversion criteria RD-to-LPP
Change non-zero color function to zero
The criteria to determine whether a droplet should be repre- Replace perturbed flow field by undisturbed flow field
Conversion
Add droplet iRD into LPP array and store droplet position and
sented by the LPP model or remain resolved depend on its shape velocity
and position. end if
The present LPP force model assumes the droplet to be spheri- end for
cal. When the droplet Weber and Reynolds are large, or when the for iLPP =1,NLPP do
droplet Capillary number Cap ¼ Wep =Rep becomes large the dro- Check conversion criteria
if Droplet iLPP not qualified for LPP
plet may deform significantly and the LPP model will introduce LPP-to-RD Reconstruct color function for the droplet
an error in its motion. However, if the droplet diameter is less than Conversion Change velocity within the droplet
four to six grid spacings, then the shape and the dynamics of the Remove droplet iLPP from LPP array
end if
droplet are not accurately represented even by the VOF method.
end for
Therefore, droplets with moderate Weber and Reynolds numbers
should also be converted to LPP droplets. In previous atomization Fig. 4. Algorithm for conversation between resolved droplets (RD) and Lagrangian
simulations these poorly resolved droplets, with dp 6 4 Dx, were point-particle (LPP) droplets.
removed to improve the stability of the solver (Shinjo and
Umemura, 2010). Here, instead of simply removing them, we con-
vert them to LPP droplets, with a positive effect on the numerical =0
=0 Reconstructed
stability and on the mass conservation, and with better droplet u undisturbed u
statistics during the atomization process. flow field
For convenience of the calculation, the droplet volume V p is u p=
used instead of its diameter dp to define the droplet size. The aspect (∫u dV)/Vp
ratio c is also estimated in order to avoid converting to LPP dro- =1 RD-to-LPP
plets thin ligaments and sheets. Therefore, only when V p is smaller
than a critical value, say V crit ’ ð4DxÞ3 , and c is close to one, then
the droplet is qualified for the LPP model. Resolved Drop (a) Lag Point-Particle
Furthermore, a location criterion is used to check if the droplet
is away from the liquid jet interface. Since the current LPP model
does not include either the formation of a droplet or a droplet =0 =0
u u
impact on the interface, only droplets that are at a certain distance
away from the liquid jet interface can be converted. The distance is
typically chosen to be the droplet diameter dp . up+u’p =0
The overall conversion criteria can be expressed as =1 LPP-to-RD up

fLPP Qualifiedg ¼ fV p < V cut g && fjcp  1j < c g && fxp 2 Rai g;
ð16Þ
where c is the tolerance for the aspect ratio, and Rai is the region
away from the interface.
RD-to-LPP conversion
For the conversion from resolved droplets to LPP droplets, the
separated liquid structures must be first identified by tagging the
color function field. The tagging approach of Herrmann (2010) is
employed here, and cells that are attached to each other, with
respect to the liquid phase, will have the same tag number.
During the tagging process, calculations of the droplet properties,
including its volume, aspect ratio, location and velocity of the cen-
ter of mass, are also performed.
A schematic representation of the conversion from RD to LPP is
shown in Fig. 5(a). In the conversion, the droplet volume V p and
velocity up are first computed by an integration over the region
X occupied by the droplet (gray area of Fig. 5(a)), namely
R R
V p ¼ X v dV and V p up ¼ X u v dV, the phase-characteristic func-
tion v is then changed from the value 1 to 0.
Moreover, the perturbed flow field in the same region should be
replaced by the undisturbed flow, so that the LPP droplet sees the
proper undisturbed fluid velocity. This is done by interpolating
each component of the fluid velocity along the corresponding coor-
dinate from the surface into the interior of a cubic region centered
at the particle location. The edge length of the cubic region is about
Resolved Drop (b) Lag Point-Particle
Fig. 5. Conversion between a resolved droplet (RD) and a Lagrangian point-particle
(LPP) droplet. The fluid flow field for the LPP droplet is slightly perturbed due to the
coupling force.
two times the droplet diameter (blue1 area of Fig. 5(a)). For exam-
ple, for the velocity component u along the x-direction, a linear
interpolation is performed between the two values uði1  1; j; kÞ
and uði2 ; j; kÞ, where i1 and i2 are the indexes of the left and right
boundary cells of the cubic region. The reconstructed velocity field
is globally divergence-free if the velocity on the surface of the
interpolation region is divergence-free. The field is also
divergence-free locally if the divergence components
@u=@x; @ v =@y, and @w=@z are constant within the region. Unless
the particle Reynolds number is very small, a cubic region with
an edge size twice the droplet diameter is sufficient to reconstruct
the undisturbed flow field in a satisfactory way. The reconstruction
of the undisturbed flow field is necessary even for the simple
approach of removing the small poorly-resolved droplets from
1
For interpretation of color in Figs. 5, 10, 11, 14 and 16, the reader is referred to the
web version of this article.
 Y. Ling et al. / International Journal of Multiphase Flow 76 (2015) 122–143
the simulation, since which minimizes the influence of the removal
of droplets to the resolved flows. It has been observed in the previ-
ous work that if the small liquid structures are replaced by gas
with the same velocity, the interface instability in the resolved
flow is modified (Shinjo and Umemura, 2010).
Since in the RD-to-LPP conversion, the region originally occu-
pied by the liquid will be filled with the gas phase, it might seem
that additional gas is added to the system and mass conservation
is violated. As a matter of fact, the added gas is actually considered
as a ‘‘ghost fluid’’ and is used only for the droplet force calculation.
When an LPP droplet is converted back to a resolved droplet, the
ghost fluid will be taken out from the system. Therefore, the ghost
fluid should always be excluded from the calculation of the total
mass of the system.
Furthermore, the LPP droplet represents not only the momen-
tum of the droplet, but also the momentum of the unresolved per-
turbed flow around the droplet. The latter can be viewed as the
momentum of a virtual volume of fluid that moves with the LPP
droplet. Therefore, although the perturbed flow around the droplet
is removed in the RD-to-LPP conversion, the corresponding
momentum is kept. As a result, the momentum of the overall sys-
tem is conserved.
LPP-to-RD conversion
Several conversion criteria can be devised to examine if an LPP
droplet is still qualified for the LPP model, for example the droplet
volume can increase due to collision and merging of droplets or to
phase change. These physical processes are not considered in the
present LPP model, where the only condition for converting an
LPP droplet to a resolved one is that the droplet position is very
close to the liquid jet interface.
A schematic picture of the conversion process is shown in
Fig. 5(b), and the algorithm is described in Fig. 4. First, a spherical
droplet is rebuilt around xp , by updating properly the color func-
tion in the cells that will be occupied by the resolved droplet.
The ghost fluid that was added in the inverse RD-to-LPP conversion
is now taken out.
In the same cells containing the new resolved droplet the veloc-
ity field needs to be updated to take care of the LPP droplet veloc-
ity. In principle, the local perturbed flow inside and outside the
droplet should be reconstructed, this being the inverse process of
the removal of the perturbed flow in the RD-to-LPP conversion.
The reconstruction of the local perturbed flow is quite a difficult
task, except in the limit of zero Reynolds number for which an ana-
lytical solution exists (however, in that case the perturbed region is
infinitely large). Nevertheless, if one naively ignores the effect of
the local perturbed flow and only changes the velocity inside the
droplet to be the LPP droplet velocity up , then it can be easily
shown that the droplet velocity will have a significant transition
jump right after the LPP-to-RD conversion. This is simply due to
the fact that a portion of the momentum of the droplet given in
the conversion is transferred to the surrounding flow and used
for the local perturbed flow development.
As described in the RD-to-LPP conversion, the LPP droplet
actually contains the momentum of both the droplet and the vir-
tual fluid moving with the droplet. In the naive conversion
approach described above, only the portion of droplet is included.
In order to compensate the perturbed flow that is not recon-
structed, the momentum of the virtual fluid also needs to be
added back. Therefore, the velocity in the cells occupied by the
resolved droplet should be changed to u0p for the purpose of com-
pensating the momentum in the perturbed flow field around the
droplet, i.e.,
Z size PDF and droplet flying angle PDF. To validate the combined
qp V p ðu0p  u~ g Þ ¼ qp V p ðup  u~ g Þ þ qf ðug  u~ g ÞdV;
Vp
129
The second term on the right hand side is the momentum of the
perturbed flow field, which is approximated by assuming a virtual
volume of fluid, aV p , moving with the droplet velocity,
Z
qg ðug  u~ g ÞdV  aqg V p ðup  u~ g Þ: ð18Þ
Vp
Then the velocity to be used in the cells of the resolved droplet
in the LPP-to-RD conversion can be expressed as
!
qg
u0p  ug ¼ 1 þ a ðup  ug Þ: ð19Þ
qp
The physical meaning of a is the ratio between the volumes of
the virtual fluid and the droplet. However, it is not easy to get a
precise value of a. It is known that, in order to accelerate a particle,
additional forces, i.e., the added-mass and history
(viscous-unsteady) forces, are needed to accelerate the ambient
fluid around the particle. For the added-mass effect, it is considered
that the ratio between the volumes of the virtual fluid to be accel-
erated through the inviscid mechanism and the particle is the
added-mass coefficient C M , which is equal to 0.5 for a sphere in
incompressible flows. Due to finite viscous diffusion time scale,
the history force usually needs to be expressed in integral form
(Basset, 1888; Mei and Adrian, 1992). Nevertheless, it has been
shown by Ling et al. (2013) that, if the particle and ambient fluid
acceleration time scales are much larger than the viscous diffusion
time scale, the history force can also be expressed as non-integral
form like the added-mass force, and a viscous-unsteady coefficient
C v u similar to the added-mass coefficient C M can be derived as
 
0:75 þ 0:105Rep
C v u  8:51 : ð20Þ
Rep
The viscous-unsteady coefficient C v u , can be considered as the
ratio between the volumes of the virtual fluid to be accelerated
through the viscous mechanism and the particle. It is shown that
C v u decreases with the droplet Reynolds number Rep . When Rep
varies from 10 to 100, C v u decreases from 1.53 to 0.96. The excess
momentum added through u0p is to mimic the effects of the
added-mass and history forces on accelerating the surrounding
fluid around the droplet. Therefore, a can be estimated as the
sum of C M and C v u and thus depends on the droplet Reynolds num-
ber as well. In an actual computation, the resolution of RD is usu-
ally 4–6 grid cells across the diameter, and the local perturbed flow
may not be computed very accurately. As a result, a usually needs
to be adjusted slightly. For simplicity, a constant of a ¼ 4 is used in
the present study, which has ben shown to yield a smooth droplet
velocity transition in the LPP-to-RD conversion.
Results
To validate the combined multiphase flow model where the
smallest droplets are represented by the LPP model, four test prob-
lems with increasing complexity are conducted: settling droplet,
droplet in a lid-driven cavity, pulsed jet atomization, and
gas-assisted atomization. The combined model considers the
point-particle approximation for small droplets, that introduces
errors even when the droplets are spherical. Nevertheless, these
errors should be small and should not influence the features of
interest to be captured by atomization simulations, such as droplet
ð17Þ model, its numerical results are compared with experimental data
or DNS, depending on the test problem.
130 Y. Ling et al. / International Journal of Multiphase Flow 76 (2015) 122–143

Settling droplet Table 1

Parameters for the settling droplet problem.

The combined model is first tested with the simulation of a liq-
uid droplet settling in a large rectangular cuboid filled with a sta-
tionary gas, see Fig. 6(a). The relevant physical properties of the
two phases and other parameters are given in Table 1. The dimen-
sions of the tank are Lx ¼ Lz ¼ Ly =2 ¼ 103 m. Gravity is along the
negative y-direction. The center of the droplet is initially at
(0:5 Lx ; 0:5 Ly ; 0:5 Lx ) and reaches the settling velocity at about time
t ¼ 0:03 s. The droplet terminal velocity is about v t ¼ 0:04 m/s
with a corresponding droplet Reynolds number Ret  0:4. The dro-
plet remains about spherical during the settling due to viscous
effects, and surface tension is therefore not included in this
simulation.
Note that the standard drag correlation Eq. (10) is for LPP in an
unbounded domain. Here, the tank width and depth are only 10
times the droplet diameter and, as a result, the tank wall has a
noticeable retarding effect on the settling velocity v t . The relation
between the terminal settling velocity v t and the
diameter-to-width ratio dp =Lx has been studied by many research-
ers, typically in a cylindrical tube with a circular cross section (Di
Felice, 1996). To include the wall effect, the correlation of Di
Felice (1996)
 2:7
vt 1  dp =Lx
g¼ ¼ ;
v t;unbounded 1  0:33 dp =Lx
is used to correct the drag calculation, where g is defined by the
ratio between the terminal settling velocity in bounded and
unbounded domains. This correlation is here used to correct the
quasi-steady force in the equation of motion (6) in the following
way
 
dup u ~ g  up / qg Du
~ g C m qg Du~ g dup
¼ þ þ  þ g:
dt sp g qp Dt qp Dt dt
In DNS without the LPP model and with cubic cells of side h,
four different resolutions, dp =h ¼ 2; 4; 8; 16, are used to resolve
the droplet. With the LPP model there is only one droplet in the
tank, hence the one-way coupling approximation is made and f p
~ g and Du
is ignored in Eq. (1). As a result, u ~ g =Dt are equal to zero.
The LPP model result is simply obtained by integrating Eq. (22)
and thus it is independent of the cell size. The results for the
y-velocity of the droplet with only the VOF method or the LPP
model are shown in Fig. 6(b). It can be observed that the LPP result
agrees well with the DNS results obtained with the finest meshes,
ql (kg/m3) qg (kg/m3) ll (Pa s) lg (Pa s) dp (m) r (N/m) g (m/s2)
1000 10 10 3
104
10 4 0 9.8
dp =h ¼ 8; 16, when the droplet is sufficiently resolved. When the
mesh resolution is low, dp =h ¼ 2; 4, DNS yields errors of 16.2%
and 7.0% in the terminal velocity, respectively. Therefore, in this
case LPP model yields more accurate results than DNS of a poorly
resolved droplet.
Though there is no experimental data for this specific test case,
there are empirical correlation for the terminal settling velocity to
compare with. For a droplet settling in an unbounded domain, the
terminal velocity obtained through the standard drag (Clift and
Gauvin, 1970) is about v t ¼ 0:0493 m/s. This value is considerably
larger than the simulation results, implying that the wall effect is
significant and should be considered. As a function of the
Reynolds number, in the two limits Rep ! 0 and Rep ! 1, the ter-
minal velocity, including the wall retarding effect, can be computed
as v t;0 ¼ 0:0393 m/s and v t;1 ¼ 0:0478 m/s (Di Felice, 1996). Both
DNS and LPP results, with the exception of the lowest resolution
dp =h ¼ 2, are bounded by these two limits. In particular, since the
Reynolds number based on the droplet terminal velocity is about
ð21Þ Rep ¼ 0:4, we expect and find that our results are closer to the
lower limit. With the correlation of Di Felice (1996) the terminal
velocity is about v t ¼ 0:0406 m/s, which again agrees reasonably
well with our simulations. The small discrepancy may be also
due to the different shape of the domain cross section in the pre-
sent study and in Di Felice (1996).
Lid-driven cavity
ð22Þ
In the second test problem, we investigate the droplet motion in
a 3D lid-driven cavity. The cavity is a cube of size L and the flow is
driven by a moving lid on the upper boundary with velocity U 0 ; the
Reynolds number of the cavity is Rec ¼ 100. The boundary condi-
tions along the x- and y-directions are no-slip, periodic boundary
conditions are considered along the third z-direction. The relevant
physical properties of the two phases and other parameters are
given in Table 2. Simulations are first run without any droplet, then
the induced flow remains 2D, moving towards the steady state
condition shown in Fig. 7. In Fig. 7(b), the u component of the
velocity (x-direction) along the vertical centerline y ¼ z ¼ L=2,

0
LPP
DNS,dp/h= 2
DNS,dp/h= 4
-0.01 DNS,dp/h= 8
DNS,dp/h=16

-0.02
v (m/s)

-0.03

-0.04

-0.05
0 0.005 0.01 0.015 0.02 0.025 0.03
t (s)
(a) (b)
Fig. 6. A liquid droplet settling in a stationary gas tank: (a) problem setting, (b) settling velocity as a function of time.
 Y. Ling et al. / International Journal of Multiphase Flow 76 (2015) 122–143 131

Table 2 considered, up1 ¼ ð0:184 U 0 ; 0:0293 U 0 ; 0Þ and up2 ¼ ð0:0320 U 0 ;

Parameters for the 3D lid-driven cavity problem.
0:0320 U 0 ; 0:0320 U 0 Þ. The first value is equal to the local fluid
ql qg ll lg dp r (N/m) L (m) U0 velocity, up1 ¼ uðxp Þ, the second value has a different magnitude
(kg/m3) (kg/m3) (Pa s) (Pa s) (m) (m/s) and direction with respect to the local fluid velocity. With the ini-
3000 10 103 104 104 1:5  105 3:2  104 3.125 tial velocity up1 the droplet will be represented by the LPP model in
the slab defined by the two limits ymin ¼ 0:61 L and ymax ¼ 0:64 L,
with up2 the limits are ymin ¼ 0:59 L and ymax ¼ 0:625 L. The simula-
tions are performed with the mesh resolution L=h ¼ 128, corre-
and the v-component (y-direction) along the horizontal centerline
sponding to a ratio dp =h ¼ 4.
x ¼ z ¼ L=2 are compared with the classic work of Ghia et al.
The droplet is initially seeded into the driven cavity as an LPP
(1982). Good agreement is observed with the results obtained with
droplet. For the case of DNS, the LPP droplet is converted to a
the two grid resolutions L=h ¼ 64; 128.
resolved droplet right after it is seeded. To have a fair comparison
After the cavity flow reaches a stationary condition at about
between the LPP and DNS results, it is important the average veloc-
time t ¼ 1 ms, a single or multiple droplets are seeded into the cav-
ity of the resolved droplet just after the LPP-to-RD conversion
ity. We first consider the simpler case where only one droplet is
matches well with the LPP droplet velocity at the corresponding
seeded. As in the settling droplet problem, the feedback effect of
time. Different values of a are tested and the results are summa-
the single LPP droplet to the resolved flow is ignored, hence f p ¼ 0.
rized Table 3. The superscript 1 and 2 denote the droplet velocity
Four different computations are performed: DNS (RD), LPP, and
right before and 6 steps after the LPP-to-RD conversion. The dro-
two different implementation of the RD-LPP conversion. In the first
plet velocities for LPP-to-RD conversion with different a are repre-
one (RD-LPP-1), the complete conversion algorithm described in
sented by u2p;RD ðaÞ; while the LPP droplet at the corresponding time
Section ‘Two-way conversion between LPP and resolved droplet’
and Fig. 4 is used; while in the second one (RD-LPP-2) the velocity without a conversion is denoted by u2p;LPP . It can be seen that among
remains unchanged in the conversion, namely the undisturbed and the values tested, a ¼ 4 yields results closest to u2p;LPP .
disturbed flow fields are not reconstructed in the two conversions The time evolution of the droplet velocity for the two initial
LPP-to-RD and RD-to-LPP. The conversion criteria for this particu- velocities is shown in Figs. 8 and 9. First, it should be observed that
lar test case is that the droplet is represented by the LPP model the DNS and LPP results agree reasonably well for both cases, as the
inside the region ymin < yp < ymax . difference in magnitude of the u component in Fig. 8(a) is about
The initial position of the droplet is at 10%. If this value is compared to the time variation of the v compo-
xp ¼ ð0:61 L; 0:61 L; 0:59 LÞ, and two different initial velocities are nent in Fig. 8(b), then the difference between DNS and LPP is

N=128
N= 64
Ghia

(a)

(b)
Fig. 7. Single phase results of the lid-driven cavity problem: (a) velocity vector plot at z ¼ L=2, (b) u at y ¼ z ¼ L=2 and v at x ¼ z ¼ L=2.

Table 3
The effect of the paremeter a on the droplet velocity in LPP-to-RD conversion. The superscript 1 and 2 denote the droplet velocity right before and 6 steps after the LPP-to-RD
conversion. The droplet velocities for LPP-to-RD conversion with different a are represented by u2p;RD ðaÞ; while the LPP droplet at the corresponding time without a conversion is
denoted by u2p;LPP . The lid velocity U 0 ¼ 3:125 m/s and the undisturbed fluid velocity at the LPP droplet location before conversion is (0:1839U 0 ; 0:02928U 0 ; 0).

u1p;LPP u2p;LPP u2p;RD ða ¼ 0Þ u2p;RD ða ¼ 2Þ u2p;RD ða ¼ 4Þ u2p;RD ða ¼ 6Þ

u=U 0 0.03200 0.03190 0.02871 0.03013 0.03155 0.03297

v =U 0 0.03200 0.03200 0.03377 0.03380 0.03383 0.03386
w=U 0 0.03200 0.03198 0.03007 0.03027 0.03047 0.03067
132 Y. Ling et al. / International Journal of Multiphase Flow 76 (2015) 122–143

-0.2 0.7

-0.25 DNS (RD) 0.6 DNS (RD)

LPP LPP
-0.3 RD-LPP-1 0.5 RD-LPP-1
RD-LPP-2 RD-LPP-2
-0.35 0.4
u (m/s)

v (m/s)
-0.4 RD-to-LPP 0.3 RD-to-LPP

-0.45 0.2

-0.5 0.1

-0.55 0
LPP-to-RD LPP-to-RD
-0.6 -0.1
0.001 0.00105 0.0011 0.00115 0.0012 0.001 0.00105 0.0011 0.00115 0.0012
t (s) t (s)
(a) (b)
Fig. 8. Time evolution of the velocity components when the droplet initial velocity is the same as the local fluid velocity: (a) u, (b) v.

0.1 0.3
DNS (RD)
0 LPP DNS (RD)
RD-LPP-1 0.2 LPP
-0.1 RD-LPP-2 RD-LPP-1
RD-LPP-2
-0.2 0.1
v (m/s)
u (m/s)

LPP-to-RD
-0.3

-0.4 0
LPP-to-RD
-0.5
-0.1
-0.6

-0.7 -0.2
0.001 0.0011 0.0012 0.0013 0.001 0.0011 0.0012 0.0013
t (s) t (s)

(a) (b)
Fig. 9. Time evolution of the velocity components when the droplet initial velocity is different from the local fluid velocity: (a) u, (b) v.

indeed quite small. Furthermore, this error magnitude is also con-
sistent with the droplet low resolution, dp =h ¼ 4.
When we consider the conversion model with the initial veloc-
ity up1 we observe that the droplet at about time t ¼ 1 ms is on the
lower boundary of the slab with a negative y-component of the
velocity. The LPP particle is quickly converted to RD, then after a
transient it changes the sign of this velocity component and finally
it enters the slab region where it is converted back to LPP. A similar
discussion holds for the initial velocity up2 . With the conversion
model described in Section ‘Two-way conversion between LPP
and resolved droplet’, the droplet velocity has a smooth transition
after the conversions for both initial velocities up1 and up12 . In
Figs. 8 and 9 this model is represented by the RD-LPP-1 lines, with
the conversion instants marked by an arrow. If the perturbed flow
field is not reconstructed properly in the LPP-to-RD conversion,
then remarkable transition jumps are observed in the droplet
velocity after the conversion (RD-LPP-2 lines). The transition jump
is more significant when the droplet is not in velocity equilibrium
with the local fluid.
As the drag force tends to bring the droplet in velocity equilib-
rium with the fluid, the droplet velocity may slowly change toward
the correct value, as in Fig. 8, but the transient process can be quite
long. However, if the initial error in the conversion causes a signif-
icant deviation of the droplet trajectory, the droplet velocity will
never go back to the correct value, as in Fig. 9(b). Therefore, it is
critical to handle properly the local flow field in the conversion
process.
We also consider the case with multiple droplets seeded in the
cavity. In total 183 monodisperse droplets are seeded randomly
after a steady state flow is reached in the cavity with a single
phase. The mass fraction of the seeded droplets, defined by the
ratio of their mass to the total mass of liquid droplets and gas, is
about 0.75. In this case the mass fraction is large, and two-way
coupling simulations are performed with the force coupling model
described in Section ‘Momentum coupling between LPP and
resolved flow’. The size of the parameter that controls the force dis-
tribution is L ¼ 10 dp .
Fig. 10 shows the droplet distribution inside the cavity at times
t ¼ 1:04; 1:08; 1:12 ms. When the droplets are resolved by the VOF
method the interfaces are plotted as the intermediate contour level
C ¼ 0:5, and are indicated by a red color. The droplet distribution
represented by the LPP model is given by the green spheres. The
DNS and LPP results overall match well on the droplet location.
The differences, shown in the images on the third column, mainly
occur when the droplets are subject to collisions or serious defor-
mations, as near the top of the cavity where the shear is strong.
This indicates the need for a future improvement of the model to
take into account both droplet–droplet and droplet–wall interac-
tions, such as incorporating the collision model of Schmidt and
Rutland (2000).
 Y. Ling et al. / International Journal of Multiphase Flow 76 (2015) 122–143
(a)
(b)
(c)
DN S
Fig. 10. Droplet distributions in the cavity at times: (a) t ¼ 1:04  10 , (b) t ¼ 1:08  103 , and (c) t ¼ 1:12  103 s.
It is also important to point out that the results obtained with
the one-way and the two-way coupling models are indeed very
close. Therefore, the one-way coupling with the present conversion
model seems to be adequate even when the LPP droplet mass frac-
tion is quite large. Nevertheless, further validation of the coupling
model is still required in future works.
Pulsed-jet atomization
Next we consider the atomization of a dense cylindrical liquid
jet which is injected into a stagnant gas tank, to examine the capa-
bility of the present model to simulate multiphase flows with very
complex interfaces. The primary breakup of a liquid jet injected
into stagnant gas has been investigated by many previous works
in detail, see for example (Ménard et al., 2007; Lebas et al., 2009;
Shinjo and Umemura, 2010). Discussions of the primary breakup
mechanisms of the liquid jet can be found in the above reference.
In the present paper, we mainly focus on testing the present model.
The jet diameter is denoted by Djet and the tank dimensions are
Ly ¼ Lz ¼ 4 Djet and Lx ¼ 10 Djet . The injection is along the positive
x-direction. The inflow velocity u is modulated sinusoidally,
u ¼ U jet ð1 þ 0:05 sinð20 p tÞÞ, to promote the growth of primary
shear instabilities. The physical properties and parameters in
non-dimensional form are listed in Table 4, where
Rejet ¼ ql U jet Djet =ll and Wejet ¼ ql U 2jet Djet =r.
On the left and right boundaries, inflow and outflow boundary
conditions are considered, respectively, while free-slip conditions
are applied on the other four boundaries. The grid is composed
of cubic cells of side h ¼ Djet =64. Two simulations are performed
for this problem: the first one is DNS with the VOF method, the sec-
ond one with the combined method, with the small droplets
described by the LPP model. Since the mass fraction of the LPP
133
LPP DNS vs LPP
3
Table 4
Parameters for the pulsed-jet atomization problem.
ql =qg ll =lg Rejet Wejet
20 20 800 4000
droplets in the flow is small during all the simulation, and based
on the knowledge of the previous test, the one-way coupling model
is used. Since the amount of liquid injected into the domain
increases with time, the total number of droplets and the number
of those represented by LPP model also increase. The maximum
number of LPP droplets is about 6000.
The simulation of the pulsed-jet problem is conducted on the
CALMIP machine, with a typical CPU time on 160 processors of
245 min for the combined model and 253 min for DNS.
Snapshots of the interface configuration in the two simulations
at time t ¼ 9 Djet =U jet are shown in Fig. 11(a) and (b), with the VOF
interface represented in green and white for the contour level
C ¼ 0:1, in order to make small droplets clearly visible. The orange
spheres represent the liquid droplets of the LPP model, and their
radius length is such to give the actual droplet volume. The two fig-
ures are in good agreement with each other, thus confirming the
capability of the combined model to capture atomization.
The primary breakup of the pulsed liquid jet can be clearly seen
in Fig. 11. Liquid sheets are formed at the front of the liquid jet and
also on the liquid core surface. At the front, a umbrella-shape sheet
is formed due to the impingement of the liquid jet against the stag-
nant gas and the resulting lateral liquid spread. On the liquid core
surface, it is seen that the perturbation near the inlet remains
almost sinusoidal, but as it is advected downstream, the interface
folds and thin liquid sheets are formed. The sheets incline toward
the upstream direction, looking like an umbrella. As the sheets are
134 Y. Ling et al. / International Journal of Multiphase Flow 76 (2015) 122–143
Fig. 11. Snapshots of the liquid–gas interface of the pulsed-jet atomization at time t ¼ 9 Djet =U jet : (a) DNS, (b) the combined model, where the orange droplets are from the
LPP model.
stretched, they break up into thin ligaments. The ligaments in
sequence break into small droplets due to the Plateau-Rayleigh
instability (Shinjo and Umemura, 2010).
The size distribution (PDF) of the droplets formed in the
pulsed-jet atomization at different times is shown in Fig. 12.
Since more and more liquid is injected into the domain and then
atomized to small droplets as time evolves, the droplet Sauter
mean decreases with time, while the number of small droplets
increases. It can be seen that the droplet size distributions obtained
by DNS and the combined model are in good agreement. In partic-
ular, the combined model captures the peak of the PDF, while the
difference of the Sauter means with the DNS results is less than
10%. When the LPP model is introduced, the number of tiny dro-
plets is significantly higher than the DNS results without the LPP
model, see Fig. 12. The discrepancy may be due to the erroneous
dynamics of small droplets when they remain to be resolved by
VOF. Small fragments of VOF-represented liquid tend to be recon-
structed and advected erroneously, so that these small fragments
stay where they are formed they are easily swept by and merged
into the moving bulk liquid jet. If these small fragments are con-
verted to LPP, then they can be advected away from the liquid jet
and remain conserved. As a result, less small droplets are seen in
the pure VOF simulation than the combined VOF-LPP simulation.
The forward RD-to-LPP process and the backward LPP-to-RD
conversion, from the ligament breakup to the droplet merging back
in the liquid jet, are shown in Fig. 13. In particular, a ligament
formed at the edge of the sheet is breaking due to
Plateau-Rayleigh instability in Fig. 13(a). In Fig. 13(b), it can be
seen that the initially formed droplet has a rather irregular shape.
Under the action of surface tension, the droplet becomes closer to a
spherical shape in Fig. 13(c). Then, since the droplet satisfies both
the size and aspect-ratio criteria and is enough away from the jet, it
is converted from RD to LPP, as shown in Fig. 13(d). The LPP droplet
moves for a while in the gas, then driven by the recirculation gen-
erated on the inside of the umbrella-shape front, the droplet is car-
ried back to the liquid jet. When the LPP droplet is getting close to
the interface, it does not satisfy the criteria of LPP and then under-
goes a backward conversion from LPP to RD in Fig. 13(e). Finally it
is reabsorbed by the liquid jet in Fig. 13(f). The present results indi-
cate that the two-way conversion scheme is capable to capture
complex droplet dynamics in atomization.
Gas-assisted atomization
The combined model is finally applied to simulate the
gas-assisted atomization experiment conducted by Descamps
et al. (2008). This experiment is particularly interesting to the
present study, because the drop trajectories and drop flying angle
statistics are measured in the experiment and thus can be used
to validate the capability of the present combined model in accu-
rately capturing the drop dynamics. Up to the authors’ knowledge,
the experiment by Descamps et al. (2008) is the only one in the lit-
erature that provides detailed measurements of drop trajectories.
Simulation setup
The simulation setup is shown in Fig. 14. The atomizer consists
of two planar jets: on top the faster gas jet with injection velocity
U g , and on bottom the slower liquid jet with injection velocity U l .
The inlet height for both phases is the same, and is denoted by
H. As a function of this characteristic length, the width, height
and thickness of the computational domain are
Lx =8 ¼ Ly =4 ¼ Lz ¼ H, respectively. The domain is subdivided with
cubic cells of side h ¼ H=64 for the coarse mesh and h ¼ H=128
for the fine one.
At the inlet, a separator plate, indicated by the blue color in
Fig. 14, has been positioned to separate the gas and liquid flows.
The need for such a plate to accurately simulate atomization has
been shown in previous papers (Fuster et al., 2013). The width
and height of the separator plate are lx ¼ H=4 and ly ¼ H=32,
respectively. Inflow and outflow boundary conditions are applied
along the x-direction, slip conditions along the vertical y-direction,
periodic boundary conditions along the z-direction. A snapshot of
the liquid–gas interface resolved by the VOF method (in gray)
and by the LPP model (in orange) is also shown in Fig. 14.
The relevant physical properties of the two phases and other
parameters are given in Table 5. The properties of the gas and liq-
uid phases are chosen in such a way to guarantee that the impor-
tant dimensionless parameters are as close as possible to the
corresponding value of the experiment by Descamps et al. (2008).
In particular, Table 6 shows that the Reynolds and Weber num-
bers of the liquid inflow, Rel ¼ ðql U l HÞ=ll and Wel ¼ ðql U 2l HÞ=r,
and the momentum ratio between the gas and liquid inflows,
M ¼ ðqg U 2g Þ=ðql U 2l Þ, are very close to the experimental value. On
the other hand, while the numerical Reynolds number based on
the gas inflow boundary layer thickness dg , Reg;d ¼ ðqg U g dg Þ=lg ,
is also very similar, the corresponding Weber number,
Weg;d ¼ ðqg U 2g dg Þ=r, is about twice as big. Finally, the numerical
Reynolds number of the gas jet is fairly large,
Reg ¼ ðqg U g HÞ=lg  8000, but still only about 42% of the experi-
mental value. For such high Reynolds numbers, it is extremely
expensive to resolve all turbulent scales in the gas flow, and we
do not expect that the current mesh resolution can capture the
smallest turbulent eddies. Therefore, the term ‘‘DNS’’ here refers
 Y. Ling et al. / International Journal of Multiphase Flow 76 (2015) 122–143 135

10 3 10 3 10 3
DNS DNS DNS
Model Model Model

10 2 10 2 10 2
counts

counts

counts
10 1 10 1 10 1

d32,DNS=2.60E-3 d32,DNS=2.55E-3 d32,DNS=2.33E-3

d32,LPP=2.88E-3 d32,LPP=2.66E-3 d32,LPP=2.49E-3

10 0 10 0 10 0
0 0.005 0.01 0.015 0.02 0 0.005 0.01 0.015 0.02 0 0.005 0.01 0.015 0.02
d d d
(a) (b) (c)
Fig. 12. Size distribution and Sauter mean of droplets formed in pulsed-jet atomization at times (time units Djet =U jet ): (a) t ¼ 8; d32;DNS ¼ 2:60 mm; d32;DNS ¼ 2:88 mm, (b)
t ¼ 8:5; d32;DNS ¼ 2:55 mm; d32;DNS ¼ 2:66 mm, (c) t ¼ 9; d32;DNS ¼ 2:33 mm; d32;DNS ¼ 2:49 mm.

(a) (b) (c)

(d ) (e ) (f)

Fig. 13. The two-way RD-to-LPP and LPP-to-RD conversions for a droplet initially formed from a ligament breakup and later reabsorbed by the liquid jet, at times (time units
Djet =U jet ): (a) t ¼ 2:0, (b) t ¼ 2:2, (c) t ¼ 2:4, (d) t ¼ 2:6, (e) t ¼ 2:8, (f) t ¼ 3:2.

Table 5
Parameters for the gas-assisted atomization problem.
Lz=H
Lx=8H
Periodic ql qg ll lg r Ul Ug H dg (m)
(kg/m3) (kg/m3) (Pa s) (Pa s) (N/m) (m/s) (m/s) (m)
c=0 Slip wall
1000 10 103 104 0.2 8.3 0.1 0.01 0.0011
Inflow LPP
Separator plate
Gas H interface Outflow

Ly=4H to simulations solving the Navier–Stokes equation without the LPP

Liquid H c=1 model, but it does not necessarily mean that every physical scale is
y fully resolved. In terms of turbulence, the numerical simulation can
be considered as implicit LES instead of true ‘‘DNS’’. Nevertheless,
Slip wall
x the mesh resolution considered in this test seems to be sufficient
z
to capture the macro-scale features of the atomization process
Fig. 14. Problem description of gas-assisted atomization. and droplet dynamics, as shown later.
136 Y. Ling et al. / International Journal of Multiphase Flow 76 (2015) 122–143
Table 6
Important dimensionless parameters of the present simulation and the experiment by Descamps et al. (2008) of gas-assisted atomization.
ql =qg ll =lg Rel
Experiment 833 58.8 1000
Simulation 100 10 1000
The simulation starts at time t 0 ¼ 0 with no liquid inside the
domain and it takes a long transient for the liquid jet to flow into
the domain and reach an ‘‘equilibrium’’ state between the inflow
and outflow of the liquid phase. The volume fraction of liquid
within the whole domain as a function of time is plotted in
Fig. 15. It can be seen that the amount of liquid in the domain
increases gradually and finally reaches a plateau at about
te ¼ 0:4 s. To save computational time only DNS is performed up
to time t e . After reaching the ‘‘equilibrium’’ state, both DNS and
the combined model are executed up to the final time tf ¼ 0:6 s.
When running the combined model, droplets with an equivalent
diameter smaller than four grid spacings are converted to
Lagrangian particles if all conversion criteria are satisfied. The total
number LPP droplets varies in time, the maximum value can reach
up to 60,000. The simulations with a grid resolution h ¼ H=64, are
conducted on 256 processors of the CINES-JADE machine: about
33 h of CPU time are required to reach time t e , and 12.5 more hours
to time tf , either with DNS or the combined model.
General behavior
The liquid jet atomization assisted by the fast gas stream is
shown in Fig. 16 at different times. The gray and orange colors indi-
cate the interface resolved by the VOF method and the small dro-
plets represented by the LPP model, respectively. The vector plot
on the background shows the flow velocity, where the vector mag-
nitude is indicated by both length and color of the arrow. The flow
velocity within the liquid is much smaller than in the gas, and the
length of the arrows is too small to be appreciated. The simulation
results are shown in Fig. 16 on a xy-plane. Due to the large velocity
difference between the liquid and gas jets, a Kelvin–Helmholtz
instability develops when the two streams meet at the end of the
separator plate (Marmottant and Villermaux, 2004; Fuster et al.,
2013). The interface wave appearing near the inlet, which is indi-
cated by a purple arrow, is at first characterized by a small ampli-
tude and by a shape close to a sinusoidal wave, see Fig. 16(a). As
the wave propagates downstream, its amplitude grows and the
wave crest rises into the fast gas stream, introducing a significant
perturbation into the gas stream that strips small droplets away
from the wave, see Fig. 16(b). Since the wave propagates much
slower than the gas stream, it acts like an obstacle to the gas flow.
As a consequence, the gas flow over the wave separates and forms
a recirculation zone, indicated by a yellow arrow in Fig. 16. The
recirculation zone formed at the downstream of the interface wave
has also been observed experimentally and numerically by Jerome
et al. (2013). The interaction between the wave and the gas stream
becomes more complex as the wave further grows and propagates
downstream, and the wave crest breaks up in a more pronounced
way and generates a large amount of droplets, as shown in
Fig. 16(b)–(d). The interface then starts to fold and liquid ligaments
are formed, see Fig. 16(e). The gas stream is significantly influenced
by the deformed interface and is observed to bend upward. A coun-
terclockwise circulation is formed near the top of the gas stream,
indicated by blue arrows Fig. 16(a), (b) and (e), (f). The clockwise
circulation at the downstream of the wave grows in time and
pushed downstream as the wave moves. These circulation regions
are observed to have a significant impact on droplet formation
(Jerome et al., 2013). The ligaments, stretched by the gas stream,
finally break up into many droplets due to a Plateau-Rayleigh
Wel M Reg;d Weg;d Reg
1.45 69.1 960 5.68 1:69  104
1.45 68.9 913 11.0 8:30  103
0.14
0.12
Liquid Volume Fraction
0.1
0.08
0.06
0.04
0.02
0
0 0.1 0.2 0.3 0.4 0.5
t (s)
Fig. 15. Time evolution of the volume fraction of liquid within the whole domain.
instability, as shown in Fig. 16(f). Finally, the atomized droplets
are advected downstream and leave the computational domain.
Droplet dynamics and flying angle statistics
Due to the strong interaction between the wavy interface and
the gas stream, the formation and dynamics of the droplets are
rather complex. The droplets are generated along the streamwise
direction with different formation mechanisms. Near the inlet, dro-
plets are mainly stripped off the interface by the gas stream, as
highlighted by the purple arrow in Fig. 16(a)–(c). Further down-
stream, the droplet formation is mainly due to ligament
break-up, or a combination of the two mechanisms, see
Fig. 16(d)–(f).
Droplets are not only generated at different streamwise posi-
tions by different mechanisms, but their dynamics changes
remarkably as well, as shown in Fig. 17 by the trajectories of the
droplets formed near the two streamwise positions x1 ¼ 2 H and
x2 ¼ 4 H, at times t1 ¼ 0:46 s and t 2 ¼ 0:49 s, respectively. In the
figure are also plotted a snapshot of the interface and the stream-
wise velocity component, on the plane z ¼ 0, to show the influence
of the gas-interface interaction on the droplet dynamics.
In particular in Fig. 17(a) at time t1 , droplets are formed at the
wave crest when its amplitude is still relatively small and the
interaction with the gas stream is not very important. As a result,
the gas stream remains fairly horizontal and the flying angle of
the droplet is generally small (a flying angle equal to 0 is along
the x-direction). Except for a few large droplets that cross the
whole gas stream, most of them stay in its interior as they are con-
vected downstream.
The flying angle statistics for the droplets in the sampling rect-
angular cuboid A of Fig. 17(a), and defined by the coordinates
ranges 4:5 H 6 x  5:5 H; H 6 y 6 3 H, and 0 6 z 6 H, is shown in
Fig. 18. The time interval for the sampling is 0:45 6 t 6 0:47 s,
which is chosen to select the droplets that are formed close to
the inlet. From Fig. 17(a), it can be observed that the droplet trajec-
tories are quite straight after the early acceleration within the gas
stream. Therefore, it is expected that the flying angles vary little
while the droplets are moving within the sampling region. It can
 Y. Ling et al. / International Journal of Multiphase Flow 76 (2015) 122–143 137

(a) t=0.44s (d) t=0.47s

(b) t=0.45s (e) t=0.48s

(c) t=0.46s (f) t=0.49s

Fig. 16. Numerical results with the combined model of the atomizing liquid jet for times in the range 0.44–0.49 s, viewed along the z-direction. The gray and orange colors
indicate the interface resolved by VOF and the small droplets represented by LPP model, respectively. The vector plot on the background shows the flow velocity, with its
magnitude indicated by both length and color of the arrows. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of
this article.)

2D-view 3D-view
(a) t=0.46 s

(b) t=0.49 s 2D-view 3D-view

Fig. 17. 2D and 3D views of trajectories of the droplets formed at the top of the perturbation wave on the liquid–gas interface at times: (a) t ¼ 0:46 s, (b) t ¼ 0:49 s. The
interface and streamwise fluid velocity are also shown in the background.

be clearly seen in Fig. 18(a) that most of the droplets are evenly
distributed with a flying angle ranging from 30 to 30 degrees.
The statistics include all droplets, hence represented by either
the VOF method or the LPP model.
On the other hand, in Fig. 17(b) at time t 2 the wave crest is near
x2 , the interface has deformed seriously causing the gas stream to
bend upward. The liquid ligaments are mainly aligned with the
stream, so when they break-up the droplets are ejected into the
bended gas stream. As a result, most of the droplets are ejected
with a positive angle and the corresponding flying angle distribu-
tion is shifted to the right as shown in Fig. 18(b). The sampling is
done in the same volume A, but within the time interval
0:48 6 t 6 0:5 s, which mainly captures the droplets formed by
the ligaments break-up, as shown in Fig. 17(b). Because of that,
138 Y. Ling et al. / International Journal of Multiphase Flow 76 (2015) 122–143
0.5
t=4.5-4.7ss
t=0.45-0.47
0.4
N/sum (N)
0.3
0.2
0.1
0 0
-90 -60 -30 0 30
Flying Angle (deg)
(a)
Fig. 18. Numerical results of droplet flying angle distribution for the two time intervals: (a) 0:45 < t < 0:47 s, (b) 0:48 < t < 0:50 s. The sampling volume is defined by
4:5H < x < 5:5H; H < y < 3H, and 0 < z < H, and is also indicated by the letter A and the dotted yellow line in Fig. 17(a). (For interpretation of the references to colour in this
figure legend, the reader is referred to the web version of this article.)
the droplets are more aligned with the gas stream, their trajecto-
ries are less curved than for the droplets of Fig. 18(a) and their fly-
ing angle distribution in Fig. 18(b) is more peaked.
The flying angle statistics for the droplets in the sampling rect-
angular cuboid B of Fig. 17(b), and defined by the coordinates
ranges 4:5 H 6 x  5:5 H; 2:5 H 6 y 6 3:5 H and for the longer time
interval 0:4 6 t 6 0:5 s, is shown in Fig. 19. The sampling volume
B is similar to that of the experiment of Descamps et al. (2008)
and the time period is long enough to cover the overall cycle of
wave development and break up, as in Fig. 16. The simulation
results agree fairly well with the experimental data, as most of
the droplets are released with a positive flying angle and the peak
of the distribution is at about 20 degrees. However, the experimen-
tal distribution profile is narrower than the numerical one.
Droplet Reynolds and Weber number statistics
When the LPP model is applied, the relative velocity between
the droplet and the surrounding fluid can be easily obtained, being
the term u ~  up , to compute the corresponding Reynolds and
Weber numbers. The probability density function (PDF) and
mass-weighted probability density function (mPDF) of the LPP dro-
plets as a function of the Reynolds and Weber numbers are shown
in Fig. 20, for the sampling volume A of Fig. 17(a) and the two time
intervals 0:45 6 t 6 0:47 s and 0:48 6 t 6 0:5 s. In the first time
interval about 45% of the total number of droplets is represented
by the LPP model, and about 64% in the second one.
In general, the results for the two time intervals are rather sim-
ilar, even if slightly more droplets are observed in the PDF of the
second time interval at small Reynolds and Weber numbers. Both
PDFs decrease with Rep and while most of the droplets are in a
regime of very small Reynolds numbers, Rep < 1, still there is a sig-
nificant amount of droplets in the flow at higher numbers, in the
range 1 < Rep < 200 as shown in Fig. 20(a). If the statistics is made
on mass rather than droplet number, the mass-weighted PDF is
obtained. From Fig. 20(c) a large fraction of the total mass is in
the droplets with 20 < Rep < 70. It should be reminded that the
criterion dp < 4Dx is used for RD-to-LPP conversion. For this range
of Reynolds numbers, it is expected that the droplets would not
have been resolved accurately by VOF, had they remained RD.
The LPP model is actually a better alternative for these poorly
resolved droplets with a relatively large Reynolds number.
Both PDF and mPDF decrease monotonically as a function of the
Weber number, Wep , and for most droplets it is indeed quite small,
0.5
t=4.8-5.0ss
t=0.48-0.50
0.4
N/sum (N)
0.3
0.2
0.1
60 90 -90 -60 -30 0 30 60 90
Flying Angle (deg)
(b)
Wep < 3, in terms of both mass and particle number. The droplet
deformation is then expected to be small, and the assumption of
a spherical shape in the LPP model is indeed rather good.
The LPP droplet Reynolds number is plotted in Fig. 21 as a func-
tion of the ratio between the droplet and cell sizes at two different
time intervals. The scaling estimate of the relationship between the
Reynolds number and the diameter of solid particles by
Balachandar (2009) can be applied to the droplets here. When
the droplet response time sp is larger than the largest ambient flow
time scale, the relative velocity is mainly controlled by the ambient
flow velocity and can be considered as independent of dp . In this
limit it can be shown that Rep dp . On the other hand, when sp
is smaller than the smallest ambient flow time scale, such as the
Kolmogorov time scale, then the relative velocity can be estimated
by the Equilibrium Eulerian Approach as ju ~  up j  sp Du
~ =Dt, where
Du~ =Dt is the acceleration of the ambient flow and is again indepen-
dent of dp (Ferry and Balachandar, 2001). Then it can be easily
3
shown that in this limit, Rep dp . It is observed from Fig. 21(a)
and (b) that, the variation of the droplet Reynolds number with
the droplet diameter lies in between these two limits, for both
0:45 < t < 0:47 s and 0:48 < t < 0:5 s, although more droplets with
smaller Reynolds number are formed for the latter time period.
Furthermore, the droplets with larger Reynolds number agree bet-
ter with Rep dp ; while the droplets with smaller Reynolds num-
ber follow the Rep dp scaling. At last, Fig. 21 also shows that there
are a significant amount of droplets with large Reynolds number
and small diameter-to-cell-size ratio formed in atomization. It is
expected the dynamics of these droplets would be incorrect if they
remain to be resolved by VOF with low resolution (dp =h K 4). This
confirms it is necessary to employ LPP model to represent these
droplets.
Comparison of droplet trajectories with experiment
The trajectories of a sample of large droplets generated by the
numerical simulation of the atomization process are shown in
Fig. 22(a), while in the experiment the droplet trajectories are
traced through high-speed cameras (Descamps et al., 2008). Due
to the limited size of the computational domain, the comparison
between numerical and experimental trajectories can be done only
for a few droplets. The traced trajectories are characterized by a
large positive flying angle, and they are likely to correspond to dro-
plets with a big inertia. As a matter of fact, from the simulation
results we can compute trajectories of large droplets, with
 Y. Ling et al. / International Journal of Multiphase Flow 76 (2015) 122–143 139

0.5 0.5
Exp Model
t=0.4-0.5 s
0.4 0.4

N/sum (N)

N/sum (N)
0.3 0.3

0.2 0.2

0.1 0.1

0 0
-30 0 30 60 90 -30 0 30 60 90
Flying Angle (deg) Flying Angle (deg)

(a) (b)
Fig. 19. Droplet flying angle distribution: (a) experimental data, (b) numerical results. The sampling region is volume B of Fig. 17(b) with 4:5H < x < 5:5H; 2:5 < y < 3:5H for
the time interval 0:4 < t < 0:5 s. The sampling region is similar to the experimental one, with 4:5H < x < 5:5H and 2:5 < y < 4H (Descamps et al., 2008).

10 0 10 2
t=0.45-0.47s t=0.45-0.47s
t=0.48-0.50s t=0.48-0.50s
10 1
10 -1

10 0
10 -2
PDF

PDF

10 -1
10 -3
10 -2

10 -4
10 -3

10 -5 10 -4
0 50 100 150 200 0 0.5 1 1.5 2 2.5 3
Rep Wep

(a) (b)
10 -1 10 1
t=0.45-0.47s t=0.45-0.47s
t=0.48-0.50s t=0.48-0.50s

10 -2 10 0
mPDF

mPDF

10 -3 10 -1

10 -4 10 -2
0 50 100 150 200 0 0.5 1 1.5 2 2.5 3
Re p We p

(c) (d)
Fig. 20. Probability density function (PDF) and mass-weighted probability density function (mPDF) as a function of the Reynolds and Weber numbers, for the droplets in the
sampling volume A of Fig. 17(a), with 4:5H < x < 5:5H; H < y < 3H, and 0 < z < H, in the two time intervals 0:45 < t < 0:47 s and 0:48 < t < 0:5 s.

dp ¼ 0:24—0:5 mm, which match quite well the experimental mea-
surements. Both the numerical and experimental trajectories are
rather straight. The computed trajectories for the two droplets
with diameter dp equal to 0.24 and 0.31 mm, curve slightly as
the droplets enters the faster gas stream and are subject to a sud-
den acceleration in the streamwise direction.
The corresponding Reynolds and Weber numbers along the dro-
plets trajectory are shown in Fig. 22(b)–(c). For the largest droplet
with diameter dp ¼ 0:49 mm, which belongs to the tail of the PDF
of Fig. 20, the Reynolds number can be as high as 400, while the
Weber number remains smaller than 5. These two numbers vary
in a remarkable way as the droplets move downstream, though
140 Y. Ling et al. / International Journal of Multiphase Flow 76 (2015) 122–143

t=0.45-0.47s Rep~dp t=0.48-0.50s Rep~dp

10 2 10 2

Rep~dp3
Re p

Re p
10 1 10 1 Rep~dp3

10 0 10 0
1.00 2.00 4.00 1.00 2.00 4.00
dp/h dp/h
(a) (b)
Fig. 21. Droplet Reynolds number as a function of the ratio between the droplet and cell sizes for the droplets in the sampling volume A of Fig. 17(a), with
4:5H < x < 5:5H; H < y < 3H, and 0 < z < H, in the two time intervals 0:45 < t < 0:47 s and 0:48 < t < 0:5 s.

0.04

0.03
y (m)

0.02 dp
dp=0.24mm
Gas Jet dp
dp=0.28mm
dp
dp=0.31mm
0.01 dp
dp=0.49mm
Liquid Jet Exp
0
0 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08
x (m)
(a)
600 6
ddp=0.24mm
p dp
dp=0.24mm
500 ddp=0.28mm
p dp
5 dp=0.28mm
ddp=0.31mm
p dp
dp=0.31mm
400 ddp=0.49mm
p dp
4 dp=0.49mm
Wep
Rep

300 3

200 2

100 1

0 0
0 0.02 0.04 0.06 0.08 0 0.02 0.04 0.06 0.08
x (m) x (m)

(b) (c)
Fig. 22. Numerical results of a sample of large droplets as a function of the longitudinal x-direction: (a) projection of the trajectories on the xy-plane, (b) Reynolds number,
and (c) Weber number. The numerical trajectories are also compared with the experimental measurements.

the trajectories are quite smooth. These fluctuations are related to
the interaction with the gas turbulent flow. The turbulence of the
gas flows is shown in background of Fig. 17. However, the inertia
of these droplets is quite large, and the influence of turbulence
on their trajectory is not profound.
Droplet size distribution
Finally, the size distribution for the droplets formed in the
gas-assisted atomization is shown in Fig. 23. The droplet size dis-
tributions computed by the combined model and DNS are shown
in Fig. 23(a) and (b), respectively. The comparison of the probabil-
ity density functions (PDF) of the droplet size for the combined
model and DNS on the same size mesh is also shown in Fig. 23(d).
The results of the combined model for the droplets larger than
the grid size agree very well with the DNS result on the same size
mesh. This indicates that the LPP coupling and conversion model
do not introduce error on the droplet formation, which is mainly
resolved by VOF. When the LPP model is introduced, a larger num-
ber of droplets smaller than the cell size are observed, see
Fig. 23(a). The number of smaller-than-cell-size droplets is
 Y. Ling et al. / International Journal of Multiphase Flow 76 (2015) 122–143 141

Model, h=H/64 DNS, h=H/64

4
10 4 10

10 3 10 3
counts

counts
10 2 10 2

10 1 10 1

10 0 10 0
0 0.5 1 1.5 2 2.5 0 0.5 1 1.5 2 2.5
d (mm) d (mm)

(a) (b)

10 2
DNS, h=H/128 Model, h=H/64
DNS, h=H/64
10 5 DNS, h=H/128
10 1

10 4
counts

PDF

10 0
10 3

10 -1
10 2

10 1 10 -2
0 0.5 1 1.5 2 2.5 0 0.05 0.1 0.15 0.2 0.25
d (mm) d/H
(c) (d)
Fig. 23. Size distributions for droplets formed in atomization. (a) Counts of droplets for present model on the mesh h ¼ H=64 ¼ 0:156 mm, (b) Counts of droplets for DNS on
the mesh h ¼ H=64 ¼ 0:156 mm, (c) Counts of droplets for DNS on the mesh h ¼ H=128 ¼ 0:078 mm, (d) Probability density function (PDF) (only droplets that are larger than
the cell size are considered in PDF calculation.) The characteristic length scales for the original and finer meshes are k1 ¼ 0:166 mm and k2 ¼ 0:095 mm, respectively.

significantly lower for the DNS results without the LPP model as
shown in Fig. 23(b). The discrepancy may be due to the erroneous
dynamics of small droplets when they remain to be resolved by
VOF. Indeed small fragments of VOF-represented liquid tend to
be reconstructed and advected so erroneously that they do not
move at all, an effect that can be easily demonstrated but is never
discussed as the dynamics of such small fragments is considered
largely erroneous anyway. As these small fragments stay where
they are formed they are easily swept by and merged into the mov-
ing bulk liquid jet. If these small fragments are converted to LPP,
then they can be advected away from the liquid jet and remain
conserved. Since the size distributions for both the combined
model and DNS are possibly erroneous for the size range dp < h.
The droplets smaller than the cell size are thus excluded in the
PDF statistics shown in Fig. 23(d).
Furthermore, an exponential decay of droplet number with dro-
plet diameter for dp 2 ½0:2; 1:1 mm is observed in both Fig. 23(a)
and (b). The decay rate can be characterized by a length scale k1 ,
which is considered as the ‘‘average ligament size’’ in
Marmottant and Villermaux (2004). A least square fit of the data
of the present model for dp 2 ½0:2; 1:1 mm yields k1 ¼ 0:166 mm.
The exponential decay with k1 ¼ 0:166 is shown to match very
well with the DNS results on the same size mesh, see Fig. 23(b).
It is noted that k1 is comparable to the cell size h ¼ 0:156 mm.
In order to see the effect of grid resolution a DNS simulation is
conducted with a finer mesh, h ¼ H=128. The total number cells is
about 67 million and it cost about 400 h on 2048 cores to reach
similar physical time t ¼ 0:6 s. The detailed results of the
finer-mesh DNS are not the focus of the present paper and discus-
sion of which will be relegated to future work.
The droplet size distribution obtained by DNS on a finer mesh is
shown in Fig. 23(c). The comparison of the PDF of the droplet size
with the two different meshes is shown in Fig. 23(d). With similar
sampling time period (about 0.03 s), the droplet number for the
finer mesh is an order of magnitude larger than that for the original
mesh. It is also seen that the size distribution of small droplets
dp < 0:2 mm is slightly better resolved, but not yet converged,
when the finer mesh is used. Exponential decay of droplet number
similar to the original mesh is observed for dp 2 ½0:1; 0:6 mm in the
fine mesh results, though the decay is faster. The characteristic
length scale for the exponential decay, k2 ¼ 0:095 mm, is 42%
smaller than k1 ¼ 0:166 for the original mesh, indicating lack of
convergence, so that yet unresolved small scales in the flow are
important to droplet formation. To yield an accurate prediction
of the droplet size distribution, including the expected peak at
small scales and the exponential decay, the smaller scales must
be resolved with sufficient grid resolution. It is likely that the com-
putational cost for a fully-resolved DNS would be extremely high.
142 Y. Ling et al. / International Journal of Multiphase Flow 76 (2015) 122–143
Therefore, it may be unavoidable for future exploration of atomiza-
tion in this parameter range to develop subgrid models accounting
for the small-scale physics, such as interface instability,
interface-turbulence interaction and droplet formation at the rele-
vant scales.
Conclusions
In this paper, we introduce a multiscale simulation approach for
atomization. While the interface between different phases is
resolved by the Volume-of-Fluid (VOF) method, the small droplets
formed in atomization are represented by the Lagrangian point-
particle (LPP) model. To accurately compute the dynamics of the
LPP droplets that are larger than the grid spacing, a new model
of the momentum coupling and the two-way conversion between
the LPP droplets and the resolved flow is proposed. The combined
model is validated by a comparison with experimental data and
DNS simulations through a series of tests. A key aspect of the pre-
sent momentum coupling model is to distribute the coupling force
exerted back to the resolved flow over a length scale which is lar-
ger than the droplet diameter, say 5 to 10 times dp . By doing that
the influence of an individual LPP droplet on the local gas flow is
small and the gas flow velocity at the LPP droplet location is a close
approximation of the undisturbed gas flow velocity that is required
for the droplet force calculation. Attention is also paid on recon-
structing the local flow field around the droplet during the
two-way conversion between the resolved droplets and LPP dro-
plets. In the RD-to-LPP conversion, the undisturbed flow field is
rebuilt by interpolation from the flow properties away from the
droplet while in the reverse LPP-to-RD conversion, excess momen-
tum is added to compensate the development of the local dis-
turbed flow field. The tests clearly show that the reconstruction
of the local flow field is necessary to have a smooth transition in
droplet velocity during the conversion.
The combined VOF-LPP model is applied to simulate the
gas-assisted atomization experiment by Descamps et al. (2008).
The numerical results show complex droplet formation mecha-
nisms. The dynamics, characterized for instance by the flying angle
of droplets formed at various streamwise locations is observed to
strongly depend on the observation window. The PDF and
mass-weighted PDF of both Reynolds and Weber numbers of the
LPP droplets (with diameter smaller than four cells) are shown.
The Reynolds number of the LPP droplets changes in a wide range
(up to several hundreds), while the Weber number is in general
small. Resolving the droplets with high Reynolds number by VOF
with less than four cells will yield catastrophic errors in the predic-
tion of droplet dynamics, and it is a considerable improvement to
represent them by the LPP model. Agreement with the experimen-
tal measurements is observed for both the the flying angle distri-
bution and the trajectories of the droplets. The exponential decay
of the PDF of droplet diameters is recovered but the rate of decay
is not converged for the current grid sizes. The present model faith-
fully recovers the large scale features of atomization and the dro-
plets at scales significantly larger than the grid size, though some
small-scale physics is missed.
Acknowledgements
This project has been supported by the ANR MODEMI project
(ANR-11-MONU-0011) program, and the FIRST (Fuel Injector
Research for Sustainable Transport) project supported by the
European Commission under the 7th Framework Programme
under Grant Agreement No. 265848. The simulations of this paper
are conducted on the our lab cluster, and both the IDRIS Turing
machine and the CINES Jade machines for which we thankfully
acknowledge Grants x20142b6115 and x20142b7325 from
GENCI. The authors acknowledge Dr. G. Tryggvason and Dr. S.
Dabiri for communicating their code, which serves as the initial
version of PARIS-Simulator. We also thank T.J. Arrufat, Dr. D.
Fuster, L. Malan, and Dr. P. Yecko for their contributions to the
development of PARIS-Simulator.
References
Agbaglah, G., Delaux, S., Fuster, D., Hoepffner, J., Josserand, C., Popinet, S., Ray, P.,
Scardovelli, R., Zaleski, S., 2011. Parallel simulation of multiphase flows using
octree adaptivity and the volume-of-fluid method. C.R. Mec. 339, 194–207.
Apte, S.V., Mahesh, K., Moin, P., Oefelein, J.C., 2003. Large-eddy simulation of
swirling particle-laden flows in a coaxial-jet combustor. Int. J. Multiphase Flow
29, 1311–1331.
Apte, S.V., Gorokhovski, M., Moin, P., 2003. LES of atomizing spray with stochastic
modeling of secondary breakup. Int. J. Multiphase Flow 29, 1503–1522.
Arrufat, T.J., Dabiri, S., Fuster, D., Ling, Y., Malan, L., Scardovelli, R., Tryggvason, G.,
Yecko, P., Zaleski, S. The PARIS-Simulator code. <http://www.lmm.jussieu.fr/
zaleski/paris/index.html>.
Aulisa, E., Manservisi, S., Scardovelli, R., Zaleski, S., 2007. Interface reconstruction
with least-squares fit and split advection in three-dimensional cartesian
geometry. J. Comput. Phys. 225, 2301–2319.
Balachandar, S., 2009. A scaling analysis for point particle approaches to turbulent
multiphase flows. Int. J. Multiphase Flow 35, 801–810.
Balachandar, S., Eaton, J.K., 2010. Turbulent dispersed multiphase flow. Annu. Rev.
Fluid Mech. 42, 111–133.
Basset, A.B., 1888. On the motion of a sphere in a viscous liquid. Phil. Trans. R. Soc. A
179, 43–63.
Boeck, T., Li, J., López-Pagés, E., Yecko, P., Zaleski, S., 2007. Ligament formation in
sheared liquid–gas layers. Theor. Comp. Fluid Dyn. 21, 59–76.
Bornia, G., Cervone, A., Manservisi, S., Scardovelli, R., Zaleski, S., 2011. On the
properties and limitations of the height function method in two-dimensional
cartesian geometry. J. Comput. Phys. 230, 851–862.
Chesnel, J., Menard, T., Reveillon, J., Demoulin, F.-X., 2011a. Subgrid analysis of
liquid jet atomization. Atomization Spray 21, 41–67.
Chesnel, J., Reveillon, J., Menard, T., Demoulin, F.-X., 2011b. Large eddy simulation of
liquid jet atomization. Atomization Spray 21, 711–736.
Chorin, A.J., 1968. Numerical solution of the Navier–Stokes equations. Math.
Comput. 22, 745–762.
Clift, R., Gauvin, W.H., 1970. The motion of particles in turbulent gas streams. Proc.
Chemeca 1, 14–28.
Clift, R., Grace, J.R., Weber, M.E., 1978. Bubbles, Drops, and Particles. Dover
Publications.
Climent, E., Magnaudet, J., 2006. Dynamics of a two-dimensional upflowing mixing
layer seeded with bubbles: bubble dispersion and effect of two-way coupling.
Phys. Fluids 18, 103304.
Crowe, C.T., Sommerfield, M., Tsuji, Y., 1998. Multiphase Flows with Droplets and
Particles. CRC Press.
DeBar, R. (1974). Fundamentals of the KRAKEN code. Technical report UCIR-760,
Lawrence Livermore National Laboratory, Livermore, California, USA.
Descamps, M.N., Matas, J.-P., Cartellier, A. (2008). Gas–liquid atomisation: gas phase
characteristics by PIV measurements and spatial evolution of the spray. In:
Proceedings du 2nd colloque INCA, Initiative en Combustion Avancée.
Di Felice, R., 1996. A relationship for the wall effect on the settling velocity of a
sphere at any flow regime. Int. J. Multiphase Flow 22, 527–533.
Eaton, J.K., 2009. Two-way coupled turbulence simulations of gas-particle flows
using point-particle tracking. Int. J. Multiphase Flow 35, 792–800.
Faxén, H., 1922. Der Widerstand gegen die Bewegung einer starren Kugel in einer
zähen Flüssigkeit, die zwischen zwei parallelen ebenen Wänden eingeschlossen
ist. Ann. Der Phys. 373, 89–119.
Ferry, J., Balachandar, S., 2001. A fast Eulerian method for disperse two-phase flow.
Int. J. Multiphase Flow 27, 1199–1226.
Francois, M.M., Cummins, S.J., Dendy, E.D., Kothe, D.B., Sicilian, J.M., Williams, M.W.,
2006. A balanced-force algorithm for continuous and sharp interfacial surface
tension models within a volume tracking framework. J. Comput. Phys. 213,
141–173.
Fuster, D., Bagué, A., Boeck, T., Le Moyne, L., Leboissetier, A., Popinet, S., Ray, P.,
Scardovelli, R., Zaleski, S., 2009. Simulation of primary atomization with an
octree adaptive mesh refinement and VOF method. Int. J. Multiphase Flow 35,
550–565.
Fuster, D., Matas, J.P., Marty, S., Popinet, S., Hoepffner, J., Cartellier, A., Zaleski, S.,
2013. Instability regimes in the primary breakup instability regimes in the
primary breakup instability regimes in the primary breakup region of planar
coflowing sheets. J. Fluid Mech. 736, 150–176.
Gatignol, R., 1983. The Faxén formulae for a rigid particle in an unsteady non-
uniform Stokes flow. J. Mech. Theor. Appl. 1, 143–160.
Ghia, U., Ghia, K.N., Shin, C.T., 1982. High-Re solutions for incompressible flow using
the Navier–Stokes equations and a multigrid method. J. Comput. Phys. 48, 387–
411.
Herrmann, M., 2010. A parallel Eulerian interface tracking/lagrangian point particle
multi-scale coupling procedure. J. Comput. Phys. 229, 745–759.
 Y. Ling et al. / International Journal of Multiphase Flow 76 (2015) 122–143
Hirt, C.W., Nichols, B.D., 1981. Volume of fluid (VOF) method for the dynamics of
free boundaries. J. Comput. Phys. 39, 201–225.
Jerome, J.J.S., Marty, S., Matas, J.-P., Zaleski, S., Hoepffner, J., 2013. Vortices catapult
droplets in atomization. Phys. Fluids 25, 112109.
Josserand, C., Lemoyne, L., Troeger, R., Zaleski, S., 2005. Droplet impact on a dry
surface: triggering the splash with a small obstacle. J. Fluid Mech. 524, 47–56.
Labourasse, E., Lacanette, D., Toutant, A., Lubin, P., Vincent, S., Lebaigue, O.,
Caltagirone, J.-P., Sagaut, P., 2007. Towards large eddy simulation of isothermal
two-phase flows: governing equations and a priori tests. Int. J. Multiphase Flow
33, 1–39.
Lafaurie, B., Nardone, C., Scardovelli, R., Zaleski, S., Zanetti, G., 1994. Modelling
merging and fragmentation in multiphase flows with SURFER. J. Comput. Phys.
113, 134–147.
Lakehal, D., Labois, M., Narayanan, C., 2012. Advances in the large-eddy and
interface simulation (leis) of interfacial multiphase flows in pipes. Prog.
Comput. Fluid Dyn. 12, 153–163.
Lakehal, D., Liovic, P., 2011. Turbulence structure and interaction with steep
breaking waves. J. Fluid Mech. 674, 522–577.
Larocque, J., Vincent, S., Lacanette, D., Lubin, P., Caltagirone, J.-P., 2010. Parametric
study of LES subgrid terms in a turbulent phase separation flow. Int. J. Heat
Fluid Flow 31, 536–544.
Lasheras, J.C., Villermaux, E., Hopfinger, E.J., 1998. Break-up and atomization of a
round water jet by a high-speed annular air jet. J. Fluid Mech. 357, 351–379.
Le Chenadec, V., Pitsch, H., 2013. A monotonicity preserving sharp interface flow
solver for high density ratio two-phase flows. J. Comput. Phys. 249, 185–203.
Lebas, R., Menard, T., Beau, P.A., Berlemont, A., Demoulin, F.-X., 2009. Numerical
simulation of primary break-up and atomization: Dns and modelling study. Int.
J. Multiphase Flow 35, 247–260.
Legendre, D., Magnaudet, J., 1997. A note on the lift force on a spherical bubble or
drop in a low-Reynolds-number shear flow. Phys. Fluids 9, 3572–3574.
Leonard, B.P., 1979. A stable and accurate convective modelling procedure based on
quadratic upstream interpolation. Comput. Methods Appl. Mech. Eng. 19, 59–
98.
Ling, Y., Haselbacher, A., Balachandar, S., 2010. A numerical source of small-scale
number-density fluctuations in Eulerian–Lagrangian simulations of multiphase
flows. J. Comput. Phys. 229, 1828–1851.
Ling, Y., Parmar, M., Balachandar, S., 2013. A scaling analysis of added-mass and
history forces and their coupling in dispersed multiphase flows. Int. J.
Multiphase Flow 57, 102–114.
Lomholt, S., Maxey, M.R., 2003. Force-coupling method for particulate two-phase
flow: Stokes flow. J. Comput. Phys. 184, 381–405.
Lomholt, S., Stenum, B., Maxey, M.R., 2002. Experimental verification of the force
coupling method for particulate flows. Int. J. Multiphase Flow 28, 225–246.
Marmottant, P., Villermaux, E., 2004. On spray formation. J. Fluid Mech. 498, 73–
111.
Matas, J.-P., Marty, S., Cartellier, A., 2011. Experimental and analytical study of the
shear instability of a gas–liquid mixing layer. Phys. Fluids 23, 094112.
Maxey, M.R., Patel, B.K., 2001. Localized force representations for particles
sedimenting in Stokes flow. Int. J. Multiphase Flow 27, 1603–1626.
Maxey, M.R., Patel, B.K., Chang, E.J., Wang, L.P., 1997. Simulations of dispersed
turbulent multiphase flow. Fluid Dyn. Res. 20, 143–156.
Maxey, M.R., Riley, J.J., 1983. Equation of motion for a small rigid sphere in a
nonuniform flow. Phys. Fluids 26, 883–889.
143
Mei, R., Adrian, R.J., 1992. Flow past a sphere with an oscillation in the free-stream
velocity and unsteady drag at finite Reynolds number. J. Fluid Mech. 237, 323–
341.
Ménard, T., Tanguy, S., Berlemont, A., 2007. Coupling level set/vof/ghost fluid
methods: validation and application to 3d simulation of the primary break-up
of a liquid jet. Int. J. Multiphase Flow 33, 510–524.
Pope, S.B., 2000. Turbulent Flows. Cambridge University Press.
Popinet, S. The basilisk code. <http://basilisk.fr/>.
Popinet, S., 2003. Gerris: a tree-based adaptive solver for the incompressible euler
equations in complex geometries. J. Comput. Phys. 190, 572–600.
Popinet, S., 2009. An accurate adaptive solver for surface-tension-driven interfacial
flows. J. Comput. Phys. 228, 5838–5866.
Renardy, Y., Renardy, M., 2002. PROST: a parabolic reconstruction of surface tension
for the volume-of-fluid method. J. Comput. Phys. 183, 400–421.
Rudman, M., 1997. Volume-tracking methods for interfacial flow calculations. Int. J.
Numer. Methods Fluids 24, 671–691.
Rudman, M., 1998. A volume-tracking method for incompressible multifluid flows
with large density variations. Int. J. Numer. Methods Fluids 28, 357–378.
Scardovelli, R., Zaleski, S., 1999. Direct numerical simulation of free-surface and
interfacial flow. Annu. Rev. Fluid Mech. 31, 567–603.
Schmidt, D.P., Rutland, C.J., 2000. A new droplet collision algorithm. J. Comput. Phys.
164, 62–80.
Shinjo, J., Umemura, A., 2010. Simulation of liquid jet primary breakup: dynamics of
ligament and droplet formation. Int. J. Multiphase Flow 36, 513–532.
Sirignano, W.A., 2005. Volume averaging for the analysis of turbulent spray flows.
Int. J. Multiphase Flow 31, 675–705.
Smagorinsky, J., 1963. General circulation experiments with the primitive
equations. Monthly Weather Rev. 91, 99–164.
Sundaram, S., Collins, L.R., 1996. Numerical considerations in simulating a turbulent
suspension of finite-volume particles. J. Comput. Phys. 124, 337–350.
Sussman, M., Smereka, P., Osher, S., 1994. A level set approach for computing
solutions to incompressible two-phase flow. J. Comput. Phys. 114, 146–159.
Tomar, G., Fuster, D., Zaleski, S., Popinet, S., 2010. Multiscale simulations of primary
atomization. Comput. Fluids 39, 1864–1874.
Toutant, A., Chandesris, M., Jamet, D., Lebaigue, O., 2009a. Jump conditions for
filtered quantities at an under-resolved discontinuous interface. Part 1:
Theoretical development. Int. J. Multiphase Flow 35, 1100–1118.
Toutant, A., Chandesris, M., Jamet, D., Lebaigue, O., 2009b. Jump conditions for
filtered quantities at an under-resolved discontinuous interface. Part 2: A priori
tests. Int. J. Multiphase Flow 35, 1119–1129.
Tryggvason, G., Bunner, B., Esmaeeli, A., Juric, D., Al-Rawahi, N., Tauber, W., Han, J.,
Nas, S., Jan, Y.J., 2001. A front-tracking method for the computations of
multiphase flow. J. Comput. Phys. 169, 708–759.
Tryggvason, G., Scardovelli, R., Zaleski, S., 2011. Direct Numerical Simulations of
Gas–Liquid Multiphase Flows. Cambridge University Press.
Vincent, S., Balmigère, G., Caltagirone, J., Meillot, E., 2010. Eulerian–Lagrangian
multiscale methods for solving scalar equations-application to incompressible
two-phase flows. J. Comput. Phys. 229, 73–106.
Vincent, S., Brändle de Motta, J.C., Sarthou, A., Estivalezes, J.-L., Simonin, O., Climent,
E., 2014. A Lagrangian VOF tensorial penalty method for the DNS of resolved
particle-laden flows. J. Comput. Phys. 256, 582–614.